Ilia Shumailov

Mikel Bober-Irizar, Ilia Shumailov, Yiren Zhao et al. · deepmind

Machine learning is vulnerable to adversarial manipulation. Previous literature has demonstrated that at the training stage attackers can manipulate data and data sampling procedures to control model behaviour. A common attack goal is to plant backdoors i.e. force the victim model to learn to recognise a trigger known only by the adversary. In this paper, we introduce a new class of backdoor attacks that hide inside model architectures i.e. in the inductive bias of the functions used to train. These backdoors are simple to implement, for instance by publishing open-source code for a backdoored model architecture that others will reuse unknowingly. We demonstrate that model architectural backdoors represent a real threat and, unlike other approaches, can survive a complete re-training from scratch. We formalise the main construction principles behind architectural backdoors, such as a link between the input and the output, and describe some possible protections against them. We evaluate our attacks on computer vision benchmarks of different scales and demonstrate the underlying vulnerability is pervasive in a variety of training settings.

Cheng Zhang, Jianyi Cheng, Ilia Shumailov et al. · deepmind

The inference of Large language models (LLMs) requires immense computation and memory resources. To curtail these costs, quantisation has merged as a promising solution, but existing LLM quantisation mainly focuses on 8-bit. In this work, we explore the statistical and learning properties of the LLM layer and attribute the bottleneck of LLM quantisation to numerical scaling offsets. To address this, we adapt block quantisations for LLMs, a family of methods that share scaling factors across packed numbers. Block quantisations efficiently reduce the numerical scaling offsets solely from an arithmetic perspective, without additional treatments in the computational path. Our nearly-lossless quantised 6-bit LLMs achieve a $19\times$ higher arithmetic density and $5\times$ memory density than the float32 baseline, surpassing the prior art 8-bit quantisation by $2.5\times$ in arithmetic density and $1.2\times$ in memory density, without requiring any data calibration or re-training. We also share our insights into sub-8-bit LLM quantisation, including the mismatch between activation and weight distributions, optimal fine-tuning strategies, and a lower quantisation granularity inherent in the statistical properties of LLMs. The latter two tricks enable nearly-lossless 4-bit LLMs on downstream tasks. Our code is open-sourced.

Sahra Ghalebikesabi, Eugene Bagdasaryan, Ren Yi et al. · deepmind

Advanced AI assistants combine frontier LLMs and tool access to autonomously perform complex tasks on behalf of users. While the helpfulness of such assistants can increase dramatically with access to user information including emails and documents, this raises privacy concerns about assistants sharing inappropriate information with third parties without user supervision. To steer information-sharing assistants to behave in accordance with privacy expectations, we propose to operationalize contextual integrity (CI), a framework that equates privacy with the appropriate flow of information in a given context. In particular, we design and evaluate a number of strategies to steer assistants' information-sharing actions to be CI compliant. Our evaluation is based on a novel form filling benchmark composed of human annotations of common webform applications, and it reveals that prompting frontier LLMs to perform CI-based reasoning yields strong results.

David Glukhov, Ilia Shumailov, Yarin Gal et al. · deepmind

Large language models (LLMs) have exhibited impressive capabilities in comprehending complex instructions. However, their blind adherence to provided instructions has led to concerns regarding risks of malicious use. Existing defence mechanisms, such as model fine-tuning or output censorship using LLMs, have proven to be fallible, as LLMs can still generate problematic responses. Commonly employed censorship approaches treat the issue as a machine learning problem and rely on another LM to detect undesirable content in LLM outputs. In this paper, we present the theoretical limitations of such semantic censorship approaches. Specifically, we demonstrate that semantic censorship can be perceived as an undecidable problem, highlighting the inherent challenges in censorship that arise due to LLMs' programmatic and instruction-following capabilities. Furthermore, we argue that the challenges extend beyond semantic censorship, as knowledgeable attackers can reconstruct impermissible outputs from a collection of permissible ones. As a result, we propose that the problem of censorship needs to be reevaluated; it should be treated as a security problem which warrants the adaptation of security-based approaches to mitigate potential risks.

Yulin Zhou, Yiren Zhao, Ilia Shumailov et al. · deepmind

Current literature demonstrates that Large Language Models (LLMs) are great few-shot learners, and prompting significantly increases their performance on a range of downstream tasks in a few-shot learning setting. An attempt to automate human-led prompting followed, with some progress achieved. In particular, subsequent work demonstrates automation can outperform fine-tuning in certain K-shot learning scenarios. In this paper, we revisit techniques for automated prompting on six different downstream tasks and a larger range of K-shot learning settings. We find that automated prompting does not consistently outperform simple manual prompts. Our work suggests that, in addition to fine-tuning, manual prompts should be used as a baseline in this line of research.

Yue Gao, Ilia Shumailov, Kassem Fawaz et al. · deepmind

Defending against adversarial examples remains an open problem. A common belief is that randomness at inference increases the cost of finding adversarial inputs. An example of such a defense is to apply a random transformation to inputs prior to feeding them to the model. In this paper, we empirically and theoretically investigate such stochastic pre-processing defenses and demonstrate that they are flawed. First, we show that most stochastic defenses are weaker than previously thought; they lack sufficient randomness to withstand even standard attacks like projected gradient descent. This casts doubt on a long-held assumption that stochastic defenses invalidate attacks designed to evade deterministic defenses and force attackers to integrate the Expectation over Transformation (EOT) concept. Second, we show that stochastic defenses confront a trade-off between adversarial robustness and model invariance; they become less effective as the defended model acquires more invariance to their randomization. Future work will need to decouple these two effects. We also discuss implications and guidance for future research.

Franziska Boenisch, Adam Dziedzic, Roei Schuster et al. · deepmind, utoronto

Federated learning (FL) is a framework for users to jointly train a machine learning model. FL is promoted as a privacy-enhancing technology (PET) that provides data minimization: data never "leaves" personal devices and users share only model updates with a server (e.g., a company) coordinating the distributed training. While prior work showed that in vanilla FL a malicious server can extract users' private data from the model updates, in this work we take it further and demonstrate that a malicious server can reconstruct user data even in hardened versions of the protocol. More precisely, we propose an attack against FL protected with distributed differential privacy (DDP) and secure aggregation (SA). Our attack method is based on the introduction of sybil devices that deviate from the protocol to expose individual users' data for reconstruction by the server. The underlying root cause for the vulnerability to our attack is a power imbalance: the server orchestrates the whole protocol and users are given little guarantees about the selection of other users participating in the protocol. Moving forward, we discuss requirements for privacy guarantees in FL. We conclude that users should only participate in the protocol when they trust the server or they apply local primitives such as local DP, shifting power away from the server. Yet, the latter approaches come at significant overhead in terms of performance degradation of the trained model, making them less likely to be deployed in practice.

Anvith Thudi, Hengrui Jia, Casey Meehan et al. · deepmind

Differentially private stochastic gradient descent (DP-SGD) is the canonical approach to private deep learning. While the current privacy analysis of DP-SGD is known to be tight in some settings, several empirical results suggest that models trained on common benchmark datasets leak significantly less privacy for many datapoints. Yet, despite past attempts, a rigorous explanation for why this is the case has not been reached. Is it because there exist tighter privacy upper bounds when restricted to these dataset settings, or are our attacks not strong enough for certain datapoints? In this paper, we provide the first per-instance (i.e., ``data-dependent") DP analysis of DP-SGD. Our analysis captures the intuition that points with similar neighbors in the dataset enjoy better data-dependent privacy than outliers. Formally, this is done by modifying the per-step privacy analysis of DP-SGD to introduce a dependence on the distribution of model updates computed from a training dataset. We further develop a new composition theorem to effectively use this new per-step analysis to reason about an entire training run. Put all together, our evaluation shows that this novel DP-SGD analysis allows us to now formally show that DP-SGD leaks significantly less privacy for many datapoints (when trained on common benchmarks) than the current data-independent guarantee. This implies privacy attacks will necessarily fail against many datapoints if the adversary does not have sufficient control over the possible training datasets.

Eleanor Clifford, Ilia Shumailov, Yiren Zhao et al. · deepmind

Early backdoor attacks against machine learning set off an arms race in attack and defence development. Defences have since appeared demonstrating some ability to detect backdoors in models or even remove them. These defences work by inspecting the training data, the model, or the integrity of the training procedure. In this work, we show that backdoors can be added during compilation, circumventing any safeguards in the data preparation and model training stages. The attacker can not only insert existing weight-based backdoors during compilation, but also a new class of weight-independent backdoors, such as ImpNet. These backdoors are impossible to detect during the training or data preparation processes, because they are not yet present. Next, we demonstrate that some backdoors, including ImpNet, can only be reliably detected at the stage where they are inserted and removing them anywhere else presents a significant challenge. We conclude that ML model security requires assurance of provenance along the entire technical pipeline, including the data, model architecture, compiler, and hardware specification.

David Glukhov, Ziwen Han, Ilia Shumailov et al. · deepmind

Vulnerability of Frontier language models to misuse and jailbreaks has prompted the development of safety measures like filters and alignment training in an effort to ensure safety through robustness to adversarially crafted prompts. We assert that robustness is fundamentally insufficient for ensuring safety goals, and current defenses and evaluation methods fail to account for risks of dual-intent queries and their composition for malicious goals. To quantify these risks, we introduce a new safety evaluation framework based on impermissible information leakage of model outputs and demonstrate how our proposed question-decomposition attack can extract dangerous knowledge from a censored LLM more effectively than traditional jailbreaking. Underlying our proposed evaluation method is a novel information-theoretic threat model of inferential adversaries, distinguished from security adversaries, such as jailbreaks, in that success is measured by inferring impermissible knowledge from victim outputs as opposed to forcing explicitly impermissible outputs from the victim. Through our information-theoretic framework, we show that to ensure safety against inferential adversaries, defense mechanisms must ensure information censorship, bounding the leakage of impermissible information. However, we prove that such defenses inevitably incur a safety-utility trade-off.

Maximilian Kaufmann, Yiren Zhao, Ilia Shumailov et al. · deepmind

Neural networks are susceptible to adversarial examples-small input perturbations that cause models to fail. Adversarial training is one of the solutions that stops adversarial examples; models are exposed to attacks during training and learn to be resilient to them. Yet, such a procedure is currently expensive-it takes a long time to produce and train models with adversarial samples, and, what is worse, it occasionally fails. In this paper we demonstrate data pruning-a method for increasing adversarial training efficiency through data sub-sampling.We empirically show that data pruning leads to improvements in convergence and reliability of adversarial training, albeit with different levels of utility degradation. For example, we observe that using random sub-sampling of CIFAR10 to drop 40% of data, we lose 8% adversarial accuracy against the strongest attackers, while by using only 20% of data we lose 14% adversarial accuracy and reduce runtime by a factor of 3. Interestingly, we discover that in some settings data pruning brings benefits from both worlds-it both improves adversarial accuracy and training time.

Pranav Dahiya, Ilia Shumailov, Ross Anderson · deepmind

Randomness supports many critical functions in the field of machine learning (ML) including optimisation, data selection, privacy, and security. ML systems outsource the task of generating or harvesting randomness to the compiler, the cloud service provider or elsewhere in the toolchain. Yet there is a long history of attackers exploiting poor randomness, or even creating it -- as when the NSA put backdoors in random number generators to break cryptography. In this paper we consider whether attackers can compromise an ML system using only the randomness on which they commonly rely. We focus our effort on Randomised Smoothing, a popular approach to train certifiably robust models, and to certify specific input datapoints of an arbitrary model. We choose Randomised Smoothing since it is used for both security and safety -- to counteract adversarial examples and quantify uncertainty respectively. Under the hood, it relies on sampling Gaussian noise to explore the volume around a data point to certify that a model is not vulnerable to adversarial examples. We demonstrate an entirely novel attack, where an attacker backdoors the supplied randomness to falsely certify either an overestimate or an underestimate of robustness for up to 81 times. We demonstrate that such attacks are possible, that they require very small changes to randomness to succeed, and that they are hard to detect. As an example, we hide an attack in the random number generator and show that the randomness tests suggested by NIST fail to detect it. We advocate updating the NIST guidelines on random number testing to make them more appropriate for safety-critical and security-critical machine-learning applications.

Jason Ross Brown, Yiren Zhao, Ilia Shumailov et al. · deepmind

The Transformer is an extremely powerful and prominent deep learning architecture. In this work, we challenge the commonly held belief in deep learning that going deeper is better, and show an alternative design approach that is building wider attention Transformers. We demonstrate that wide single layer Transformer models can compete with or outperform deeper ones in a variety of Natural Language Processing (NLP) tasks when both are trained from scratch. The impact of changing the model aspect ratio on Transformers is then studied systematically. This ratio balances the number of layers and the number of attention heads per layer while keeping the total number of attention heads and all other hyperparameters constant. On average, across 4 NLP tasks and 10 attention types, single layer wide models perform 0.3% better than their deep counterparts. We show an in-depth evaluation and demonstrate how wide models require a far smaller memory footprint and can run faster on commodity hardware, in addition, these wider models are also more interpretable. For example, a single layer Transformer on the IMDb byte level text classification has 3.1x faster inference latency on a CPU than its equally accurate deeper counterpart, and is half the size. We therefore put forward wider and shallower models as a viable and desirable alternative for small models on NLP tasks, and as an important area of research for domains beyond this.

Joseph Rance, Yiren Zhao, Ilia Shumailov et al. · deepmind

Data augmentation is used extensively to improve model generalisation. However, reliance on external libraries to implement augmentation methods introduces a vulnerability into the machine learning pipeline. It is well known that backdoors can be inserted into machine learning models through serving a modified dataset to train on. Augmentation therefore presents a perfect opportunity to perform this modification without requiring an initially backdoored dataset. In this paper we present three backdoor attacks that can be covertly inserted into data augmentation. Our attacks each insert a backdoor using a different type of computer vision augmentation transform, covering simple image transforms, GAN-based augmentation, and composition-based augmentation. By inserting the backdoor using these augmentation transforms, we make our backdoors difficult to detect, while still supporting arbitrary backdoor functionality. We evaluate our attacks on a range of computer vision benchmarks and demonstrate that an attacker is able to introduce backdoors through just a malicious augmentation routine.

Avital Shafran, Ilia Shumailov, Murat A. Erdogdu et al. · deepmind

Model extraction attacks are designed to steal trained models with only query access, as is often provided through APIs that ML-as-a-Service providers offer. Machine Learning (ML) models are expensive to train, in part because data is hard to obtain, and a primary incentive for model extraction is to acquire a model while incurring less cost than training from scratch. Literature on model extraction commonly claims or presumes that the attacker is able to save on both data acquisition and labeling costs. We thoroughly evaluate this assumption and find that the attacker often does not. This is because current attacks implicitly rely on the adversary being able to sample from the victim model's data distribution. We thoroughly research factors influencing the success of model extraction. We discover that prior knowledge of the attacker, i.e., access to in-distribution data, dominates other factors like the attack policy the adversary follows to choose which queries to make to the victim model API. Our findings urge the community to redefine the adversarial goals of ME attacks as current evaluation methods misinterpret the ME performance.

Nicholas Boucher, Jenny Blessing, Ilia Shumailov et al. · deepmind

Text-based machine learning models are vulnerable to an emerging class of Unicode-based adversarial examples capable of tricking a model into misreading text with potentially disastrous effects. The primary existing defense against these attacks is to preprocess potentially malicious text inputs using optical character recognition (OCR). In theory, OCR models will ignore any malicious Unicode characters and will extract the visually correct input to be fed to the model. In this work, we show that these visual defenses fail to prevent this type of attack. We use a genetic algorithm to generate visual adversarial examples (i.e., OCR outputs) in a black-box setting, demonstrating a highly effective novel attack that substantially reduces the accuracy of OCR and other visual models. Specifically, we use the Unicode functionality of combining characters (e.g., ñ which combines the characters n and ~) to manipulate text inputs so that small visual perturbations appear when the text is displayed. We demonstrate the effectiveness of these attacks in the real world by creating adversarial examples against production models published by Meta, Microsoft, IBM, and Google. We additionally conduct a user study to establish that the model-fooling adversarial examples do not affect human comprehension of the text, showing that language models are uniquely vulnerable to this type of text attack.

Jason Ross Brown, Yiren Zhao, Ilia Shumailov et al. · deepmind

Given the wide and ever growing range of different efficient Transformer attention mechanisms, it is important to identify which attention is most effective when given a task. In this work, we are also interested in combining different attention types to build heterogeneous Transformers. We first propose a DARTS-like Neural Architecture Search (NAS) method to find the best attention for a given task, in this setup, all heads use the same attention (homogeneous models). Our results suggest that NAS is highly effective on this task, and it identifies the best attention mechanisms for IMDb byte level text classification and Listops. We then extend our framework to search for and build Transformers with multiple different attention types, and call them heterogeneous Transformers. We show that whilst these heterogeneous Transformers are better than the average homogeneous models, they cannot outperform the best. We explore the reasons why heterogeneous attention makes sense, and why it ultimately fails.

Peter Horvath, Ilia Shumailov, Lukasz Chmielewski et al. · deepmind

The multi-million dollar investment required for modern machine learning (ML) has made large ML models a prime target for theft. In response, the field of model stealing has emerged. Attacks based on physical side-channel information have shown that DNN model extraction is feasible, even on CUDA Cores in a GPU. For the first time, our work demonstrates parameter extraction on the specialized GPU's Tensor Core units, most commonly used GPU units nowadays due to their superior performance, via near-field physical side-channel attacks. Previous work targeted only the general-purpose CUDA Cores in the GPU, the functional units that have been part of the GPU since its inception. Our method is tailored to the GPU architecture to accurately estimate energy consumption and derive efficient attacks via Correlation Power Analysis (CPA). Furthermore, we provide an exploratory analysis of hyperparameter and weight leakage from LLMs in far field and demonstrate that the GPU's electromagnetic radiation leaks even 100 cm away through a glass obstacle.

Jiaqi Wang, Roei Schuster, Ilia Shumailov et al. · deepmind

When learning from sensitive data, care must be taken to ensure that training algorithms address privacy concerns. The canonical Private Aggregation of Teacher Ensembles, or PATE, computes output labels by aggregating the predictions of a (possibly distributed) collection of teacher models via a voting mechanism. The mechanism adds noise to attain a differential privacy guarantee with respect to the teachers' training data. In this work, we observe that this use of noise, which makes PATE predictions stochastic, enables new forms of leakage of sensitive information. For a given input, our adversary exploits this stochasticity to extract high-fidelity histograms of the votes submitted by the underlying teachers. From these histograms, the adversary can learn sensitive attributes of the input such as race, gender, or age. Although this attack does not directly violate the differential privacy guarantee, it clearly violates privacy norms and expectations, and would not be possible at all without the noise inserted to obtain differential privacy. In fact, counter-intuitively, the attack becomes easier as we add more noise to provide stronger differential privacy. We hope this encourages future work to consider privacy holistically rather than treat differential privacy as a panacea.

David Khachaturov, Yue Gao, Ilia Shumailov et al. · deepmind

Visual adversarial examples have so far been restricted to pixel-level image manipulations in the digital world, or have required sophisticated equipment such as 2D or 3D printers to be produced in the physical real world. We present the first ever method of generating human-producible adversarial examples for the real world that requires nothing more complicated than a marker pen. We call them $\textbf{adversarial tags}$. First, building on top of differential rendering, we demonstrate that it is possible to build potent adversarial examples with just lines. We find that by drawing just $4$ lines we can disrupt a YOLO-based model in $54.8\%$ of cases; increasing this to $9$ lines disrupts $81.8\%$ of the cases tested. Next, we devise an improved method for line placement to be invariant to human drawing error. We evaluate our system thoroughly in both digital and analogue worlds and demonstrate that our tags can be applied by untrained humans. We demonstrate the effectiveness of our method for producing real-world adversarial examples by conducting a user study where participants were asked to draw over printed images using digital equivalents as guides. We further evaluate the effectiveness of both targeted and untargeted attacks, and discuss various trade-offs and method limitations, as well as the practical and ethical implications of our work. The source code will be released publicly.

Augusto B. Corrêa, Yoav Gelberg, Luckeciano C. Melo et al. · deepmind

We show that iterative deployment of large language models (LLMs), each fine-tuned on data carefully curated by users from the previous models' deployment, can significantly change the properties of the resultant models. By testing this mechanism on various planning domains, we observe substantial improvements in planning skills, with later models displaying emergent generalization by discovering much longer plans than the initial models. We then provide theoretical analysis showing that iterative deployment effectively implements reinforcement learning (RL) training in the outer-loop (i.e. not as part of intentional model training), with an implicit reward function. The connection to RL has two important implications: first, for the field of AI safety, as the reward function entailed by repeated deployment is not defined explicitly, and could have unexpected implications to the properties of future model deployments. Second, the mechanism highlighted here can be viewed as an alternative training regime to explicit RL, relying on data curation rather than explicit rewards.

Yue Gao, Ilia Shumailov, Kassem Fawaz · deepmind

Machine Learning (ML) systems are vulnerable to adversarial examples, particularly those from query-based black-box attacks. Despite various efforts to detect and prevent such attacks, ML systems are still at risk, demanding a more comprehensive approach to security that includes logging, analyzing, and sharing evidence. While traditional security benefits from well-established practices of forensics and threat intelligence sharing, ML security has yet to find a way to profile its attackers and share information about them. In response, this paper introduces SEA, a novel ML security system to characterize black-box attacks on ML systems for forensic purposes and to facilitate human-explainable intelligence sharing. SEA leverages Hidden Markov Models to attribute the observed query sequence to known attacks. It thus understands the attack's progression rather than focusing solely on the final adversarial examples. Our evaluations reveal that SEA is effective at attack attribution, even on the second incident, and is robust to adaptive strategies designed to evade forensic analysis. SEA's explanations of the attack's behavior allow us even to fingerprint specific minor bugs in widely used attack libraries. For example, we discover that the SignOPT and Square attacks in ART v1.14 send over 50% duplicated queries. We thoroughly evaluate SEA on a variety of settings and demonstrate that it can recognize the same attack with more than 90% Top-1 and 95% Top-3 accuracy. Finally, we demonstrate how SEA generalizes to other domains like text classification.

Mark Vero, Fabian Kaczmarczyck, Ivan Petrov et al.

Honeypots are decoy systems mimicking real system components designed to defend against cyber attacks. Recently, LLMs increasingly serve as simulation backbones for honeypots. They enable defenders to construct high-interaction honeypots with low system security risks. However, LLM-powered honeypot development lacks a unified evaluation framework. Most evaluations consist of measuring response similarity on fixed commands, manual testing, or real-world deployment. These methods are often not scalable for development, reproducible across evaluations, representative of practical attacks, or adaptable to various attacker and honeypot configurations. In this work, we bridge this gap and propose Honeyval, a comprehensive evaluation framework for LLM-powered HTTP honeypots. We address the limitations of prior evaluations by grounding the honeypots in 16 backend applications, using AI hacking agents as attackers, employing two control tasks to monitor agent and honeypot capabilities across customizations, and defining clear and verifiable exploit goals for the attacker. Using Honeyval, we conduct an extensive evaluation of recent cost-efficient LLMs as HTTP honeypots. Our experiments highlight the promise of LLM-powered honeypots; they lead to substantially longer interactions with the attacker than rule-based baseline honeypots and are far less frequently detected even by frontier models, all while, on average, preserving a running cost advantage against agentic attackers. Further, we experiment with different counter-offensive honeypots configurations, and observe unique trade-offs, such as longer interactions at the cost of increased detection.

Edoardo Debenedetti, Ilia Shumailov, Tianqi Fan et al. · deepmind, eth-zurich

Large Language Models (LLMs) are increasingly deployed in agentic systems that interact with an untrusted environment. However, LLM agents are vulnerable to prompt injection attacks when handling untrusted data. In this paper we propose CaMeL, a robust defense that creates a protective system layer around the LLM, securing it even when underlying models are susceptible to attacks. To operate, CaMeL explicitly extracts the control and data flows from the (trusted) query; therefore, the untrusted data retrieved by the LLM can never impact the program flow. To further improve security, CaMeL uses a notion of a capability to prevent the exfiltration of private data over unauthorized data flows by enforcing security policies when tools are called. We demonstrate effectiveness of CaMeL by solving $77\%$ of tasks with provable security (compared to $84\%$ with an undefended system) in AgentDojo. We release CaMeL at https://github.com/google-research/camel-prompt-injection.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Harsh Chaudhari, Ethan Rathbun, Hanna Foerster et al.

Chain-of-Thought (CoT) reasoning has emerged as a powerful technique for enhancing large language models' capabilities by generating intermediate reasoning steps for complex tasks. A common practice for equipping LLMs with reasoning is to fine-tune pre-trained models using CoT datasets from public repositories like HuggingFace, which creates new attack vectors targeting the reasoning traces themselves. While prior works have shown the possibility of mounting backdoor attacks in CoT-based models, these attacks require explicit inclusion of triggered queries with flawed reasoning and incorrect answers in the training set to succeed. Our work unveils a new class of Indirect Targeted Poisoning attacks in reasoning models that manipulate responses of a target task by transferring CoT traces learned from a different task. Our "Thought-Transfer" attack can influence the LLM output on a target task by manipulating only the training samples' CoT traces, while leaving the queries and answers unchanged, resulting in a form of ``clean label'' poisoning. Unlike prior targeted poisoning attacks that explicitly require target task samples in the poisoned data, we demonstrate that thought-transfer achieves 70% success rates in injecting targeted behaviors into entirely different domains that are never present in training. Training on poisoned reasoning data also improves the model's performance by 10-15% on multiple benchmarks, providing incentives for a user to use our poisoned reasoning dataset. Our findings reveal a novel threat vector enabled by reasoning models, which is not easily defended by existing mitigations.

Hanna Foerster, Robert Mullins, Tom Blanchard et al.

AI agents are vulnerable to prompt injection attacks, where malicious content hijacks agent behavior to steal credentials or cause financial loss. The only known robust defense is architectural isolation that strictly separates trusted task planning from untrusted environment observations. However, applying this design to Computer Use Agents (CUAs) -- systems that automate tasks by viewing screens and executing actions -- presents a fundamental challenge: current agents require continuous observation of UI state to determine each action, conflicting with the isolation required for security. We resolve this tension by demonstrating that UI workflows, while dynamic, are structurally predictable. We introduce Single-Shot Planning for CUAs, where a trusted planner generates a complete execution graph with conditional branches before any observation of potentially malicious content, providing provable control flow integrity guarantees against arbitrary instruction injections. Although this architectural isolation successfully prevents instruction injections, we show that additional measures are needed to prevent Branch Steering attacks, which manipulate UI elements to trigger unintended valid paths within the plan. We evaluate our design on OSWorld, and retain up to 57% of the performance of frontier models while improving performance for smaller open-source models by up to 19%, demonstrating that rigorous security and utility can coexist in CUAs.

Sarah Meiklejohn, Hayden Blauzvern, Mihai Maruseac et al. · deepmind

Powerful machine learning (ML) models are now readily available online, which creates exciting possibilities for users who lack the deep technical expertise or substantial computing resources needed to develop them. On the other hand, this type of open ecosystem comes with many risks. In this paper, we argue that the current ecosystem for open ML models contains significant supply-chain risks, some of which have been exploited already in real attacks. These include an attacker replacing a model with something malicious (e.g., malware), or a model being trained using a vulnerable version of a framework or on restricted or poisoned data. We then explore how Sigstore, a solution designed to bring transparency to open-source software supply chains, can be used to bring transparency to open ML models, in terms of enabling model publishers to sign their models and prove properties about the datasets they use.

Cheng Zhang, Hanna Foerster, Robert D. Mullins et al. · deepmind

It is now a common business practice to buy access to large language model (LLM) inference rather than self-host, because of significant upfront hardware infrastructure and energy costs. However, as a buyer, there is no mechanism to verify the authenticity of the advertised service including the serving hardware platform, e.g. that it is actually being served using an NVIDIA H100. Furthermore, there are reports suggesting that model providers may deliver models that differ slightly from the advertised ones, often to make them run on less expensive hardware. That way, a client pays premium for a capable model access on more expensive hardware, yet ends up being served by a (potentially less capable) cheaper model on cheaper hardware. In this paper we introduce hardware and software platform inference (HSPI) -- a method for identifying the underlying GPU architecture and software stack of a (black-box) machine learning model solely based on its input-output behavior. Our method leverages the inherent differences of various GPU architectures and compilers to distinguish between different GPU types and software stacks. By analyzing the numerical patterns in the model's outputs, we propose a classification framework capable of accurately identifying the GPU used for model inference as well as the underlying software configuration. Our findings demonstrate the feasibility of inferring GPU type from black-box models. We evaluate HSPI against models served on different real hardware and find that in a white-box setting we can distinguish between different GPUs with between $83.9\%$ and $100\%$ accuracy. Even in a black-box setting we achieve results that are up to 3x higher than random guess accuracy. Our code is available at https://github.com/ChengZhang-98/HSPI.

Nicholas Boucher, Ilia Shumailov, Ross Anderson et al.

Several years of research have shown that machine-learning systems are vulnerable to adversarial examples, both in theory and in practice. Until now, such attacks have primarily targeted visual models, exploiting the gap between human and machine perception. Although text-based models have also been attacked with adversarial examples, such attacks struggled to preserve semantic meaning and indistinguishability. In this paper, we explore a large class of adversarial examples that can be used to attack text-based models in a black-box setting without making any human-perceptible visual modification to inputs. We use encoding-specific perturbations that are imperceptible to the human eye to manipulate the outputs of a wide range of Natural Language Processing (NLP) systems from neural machine-translation pipelines to web search engines. We find that with a single imperceptible encoding injection -- representing one invisible character, homoglyph, reordering, or deletion -- an attacker can significantly reduce the performance of vulnerable models, and with three injections most models can be functionally broken. Our attacks work against currently-deployed commercial systems, including those produced by Microsoft and Google, in addition to open source models published by Facebook, IBM, and HuggingFace. This novel series of attacks presents a significant threat to many language processing systems: an attacker can affect systems in a targeted manner without any assumptions about the underlying model. We conclude that text-based NLP systems require careful input sanitization, just like conventional applications, and that given such systems are now being deployed rapidly at scale, the urgent attention of architects and operators is required.

Jamie Hayes, Ilia Shumailov, Eleni Triantafillou et al. · deepmind

The high cost of model training makes it increasingly desirable to develop techniques for unlearning. These techniques seek to remove the influence of a training example without having to retrain the model from scratch. Intuitively, once a model has unlearned, an adversary that interacts with the model should no longer be able to tell whether the unlearned example was included in the model's training set or not. In the privacy literature, this is known as membership inference. In this work, we discuss adaptations of Membership Inference Attacks (MIAs) to the setting of unlearning (leading to their "U-MIA" counterparts). We propose a categorization of existing U-MIAs into "population U-MIAs", where the same attacker is instantiated for all examples, and "per-example U-MIAs", where a dedicated attacker is instantiated for each example. We show that the latter category, wherein the attacker tailors its membership prediction to each example under attack, is significantly stronger. Indeed, our results show that the commonly used U-MIAs in the unlearning literature overestimate the privacy protection afforded by existing unlearning techniques on both vision and language models. Our investigation reveals a large variance in the vulnerability of different examples to per-example U-MIAs. In fact, several unlearning algorithms lead to a reduced vulnerability for some, but not all, examples that we wish to unlearn, at the expense of increasing it for other examples. Notably, we find that the privacy protection for the remaining training examples may worsen as a consequence of unlearning. We also discuss the fundamental difficulty of equally protecting all examples using existing unlearning schemes, due to the different rates at which examples are unlearned. We demonstrate that naive attempts at tailoring unlearning stopping criteria to different examples fail to alleviate these issues.

Sierra Wyllie, Ilia Shumailov, Nicolas Papernot · deepmind

Model-induced distribution shifts (MIDS) occur as previous model outputs pollute new model training sets over generations of models. This is known as model collapse in the case of generative models, and performative prediction or unfairness feedback loops for supervised models. When a model induces a distribution shift, it also encodes its mistakes, biases, and unfairnesses into the ground truth of its data ecosystem. We introduce a framework that allows us to track multiple MIDS over many generations, finding that they can lead to loss in performance, fairness, and minoritized group representation, even in initially unbiased datasets. Despite these negative consequences, we identify how models might be used for positive, intentional, interventions in their data ecosystems, providing redress for historical discrimination through a framework called algorithmic reparation (AR). We simulate AR interventions by curating representative training batches for stochastic gradient descent to demonstrate how AR can improve upon the unfairnesses of models and data ecosystems subject to other MIDS. Our work takes an important step towards identifying, mitigating, and taking accountability for the unfair feedback loops enabled by the idea that ML systems are inherently neutral and objective.

Jamie Hayes, Marika Swanberg, Harsh Chaudhari et al. · deepmind

Large language models (LLMs) are susceptible to memorizing training data, raising concerns about the potential extraction of sensitive information at generation time. Discoverable extraction is the most common method for measuring this issue: split a training example into a prefix and suffix, then prompt the LLM with the prefix, and deem the example extractable if the LLM generates the matching suffix using greedy sampling. This definition yields a yes-or-no determination of whether extraction was successful with respect to a single query. Though efficient to compute, we show that this definition is unreliable because it does not account for non-determinism present in more realistic (non-greedy) sampling schemes, for which LLMs produce a range of outputs for the same prompt. We introduce probabilistic discoverable extraction, which, without additional cost, relaxes discoverable extraction by considering multiple queries to quantify the probability of extracting a target sequence. We evaluate our probabilistic measure across different models, sampling schemes, and training-data repetitions, and find that this measure provides more nuanced information about extraction risk compared to traditional discoverable extraction.

Chongyang Shi, Sharon Lin, Shuang Song et al. · deepmind

Gemini is increasingly used to perform tasks on behalf of users, where function-calling and tool-use capabilities enable the model to access user data. Some tools, however, require access to untrusted data introducing risk. Adversaries can embed malicious instructions in untrusted data which cause the model to deviate from the user's expectations and mishandle their data or permissions. In this report, we set out Google DeepMind's approach to evaluating the adversarial robustness of Gemini models and describe the main lessons learned from the process. We test how Gemini performs against a sophisticated adversary through an adversarial evaluation framework, which deploys a suite of adaptive attack techniques to run continuously against past, current, and future versions of Gemini. We describe how these ongoing evaluations directly help make Gemini more resilient against manipulation.

Hanna Foerster, Ilia Shumailov, Cheng Zhang et al.

Dynamic quantization emerged as a practical approach to increase the utilization and efficiency of the machine learning serving flow. Unlike static quantization, which applies quantization offline, dynamic quantization operates on tensors at run-time, adapting its parameters to the actual input data. Today's mainstream machine learning frameworks, including ML compilers and inference engines, frequently recommend dynamic quantization as an initial step for optimizing model serving. This is because dynamic quantization can significantly reduce memory usage and computational load, leading to faster token generation and improved model serving efficiency without substantial loss in model accuracy. In this paper, we reveal a critical vulnerability in dynamic quantization: an adversary can exploit such quantization strategy to steal sensitive user data placed in the same batch as the adversary's input. Our analysis demonstrates that dynamic quantization, when improperly implemented or configured, can create side channels that expose information about other inputs within the same batch. We call this phenomenon Quantamination, describing contamination from quantization. Specifically, we show that at least 4 of the most popular ML frameworks in use today either default to or can use configurations that leak data across the batch boundary. This data leakage, in theory, allows attackers to partially or even fully recover other users' batched input data, representing a serious privacy risk for existing ML serving frameworks.

Jamie Hayes, Ilia Shumailov, Christopher A. Choquette-Choo et al. · deepmind

State-of-the-art membership inference attacks (MIAs) typically require training many reference models, making it difficult to scale these attacks to large pre-trained language models (LLMs). As a result, prior research has either relied on weaker attacks that avoid training references (e.g., fine-tuning attacks), or on stronger attacks applied to small models and datasets. However, weaker attacks have been shown to be brittle and insights from strong attacks in simplified settings do not translate to today's LLMs. These challenges prompt an important question: are the limitations observed in prior work due to attack design choices, or are MIAs fundamentally ineffective on LLMs? We address this question by scaling LiRA--one of the strongest MIAs--to GPT-2 architectures ranging from 10M to 1B parameters, training references on over 20B tokens from the C4 dataset. Our results advance the understanding of MIAs on LLMs in four key ways. While (1) strong MIAs can succeed on pre-trained LLMs, (2) their effectiveness, remains limited (e.g., AUC<0.7) in practical settings. (3) Even when strong MIAs achieve better-than-random AUC, aggregate metrics can conceal substantial per-sample MIA decision instability: due to training randomness, many decisions are so unstable that they are statistically indistinguishable from a coin flip. Finally, (4) the relationship between MIA success and related LLM privacy metrics is not as straightforward as prior work has suggested.

Harry Langford, Ilia Shumailov, Yiren Zhao et al. · deepmind

While previous research backdoored neural networks by changing their parameters, recent work uncovered a more insidious threat: backdoors embedded within the definition of the network's architecture. This involves injecting common architectural components, such as activation functions and pooling layers, to subtly introduce a backdoor behavior that persists even after (full re-)training. However, the full scope and implications of architectural backdoors have remained largely unexplored. Bober-Irizar et al. [2023] introduced the first architectural backdoor; they showed how to create a backdoor for a checkerboard pattern, but never explained how to target an arbitrary trigger pattern of choice. In this work we construct an arbitrary trigger detector which can be used to backdoor an architecture with no human supervision. This leads us to revisit the concept of architecture backdoors and taxonomise them, describing 12 distinct types. To gauge the difficulty of detecting such backdoors, we conducted a user study, revealing that ML developers can only identify suspicious components in common model definitions as backdoors in 37% of cases, while they surprisingly preferred backdoored models in 33% of cases. To contextualize these results, we find that language models outperform humans at the detection of backdoors. Finally, we discuss defenses against architectural backdoors, emphasizing the need for robust and comprehensive strategies to safeguard the integrity of ML systems.

Jamie Hayes, Ilia Shumailov, Itay Yona · deepmind

Mixture of Experts (MoE) has become a key ingredient for scaling large foundation models while keeping inference costs steady. We show that expert routing strategies that have cross-batch dependencies are vulnerable to attacks. Malicious queries can be sent to a model and can affect a model's output on other benign queries if they are grouped in the same batch. We demonstrate this via a proof-of-concept attack in a toy experimental setting.

Itay Yona, Ilia Shumailov, Jamie Hayes et al. · deepmind

Large Language Models (LLMs), despite their impressive capabilities, often fail to accurately repeat a single word when prompted to, and instead output unrelated text. This unexplained failure mode represents a vulnerability, allowing even end-users to diverge models away from their intended behavior. We aim to explain the causes for this phenomenon and link it to the concept of ``attention sinks'', an emergent LLM behavior crucial for fluency, in which the initial token receives disproportionately high attention scores. Our investigation identifies the neural circuit responsible for attention sinks and shows how long repetitions disrupt this circuit. We extend this finding to other non-repeating sequences that exhibit similar circuit disruptions. To address this, we propose a targeted patch that effectively resolves the issue without negatively impacting the model's overall performance. This study provides a mechanistic explanation for an LLM vulnerability, demonstrating how interpretability can diagnose and address issues, and offering insights that pave the way for more secure and reliable models.

Milad Nasr, Nicholas Carlini, Chawin Sitawarin et al. · eth-zurich

How should we evaluate the robustness of language model defenses? Current defenses against jailbreaks and prompt injections (which aim to prevent an attacker from eliciting harmful knowledge or remotely triggering malicious actions, respectively) are typically evaluated either against a static set of harmful attack strings, or against computationally weak optimization methods that were not designed with the defense in mind. We argue that this evaluation process is flawed. Instead, we should evaluate defenses against adaptive attackers who explicitly modify their attack strategy to counter a defense's design while spending considerable resources to optimize their objective. By systematically tuning and scaling general optimization techniques-gradient descent, reinforcement learning, random search, and human-guided exploration-we bypass 12 recent defenses (based on a diverse set of techniques) with attack success rate above 90% for most; importantly, the majority of defenses originally reported near-zero attack success rates. We believe that future defense work must consider stronger attacks, such as the ones we describe, in order to make reliable and convincing claims of robustness.

Itay Yona, Ilia Shumailov, Jamie Hayes et al. · deepmind

Mixture-of-Experts (MoE) models improve the efficiency and scalability of dense language models by routing each token to a small number of experts in each layer. In this paper, we show how an adversary that can arrange for their queries to appear in the same batch of examples as a victim's queries can exploit Expert-Choice-Routing to fully disclose a victim's prompt. We successfully demonstrate the effectiveness of this attack on a two-layer Mixtral model, exploiting the tie-handling behavior of the torch.topk CUDA implementation. Our results show that we can extract the entire prompt using $O({VM}^2)$ queries (with vocabulary size $V$ and prompt length $M$) or 100 queries on average per token in the setting we consider. This is the first attack to exploit architectural flaws for the purpose of extracting user prompts, introducing a new class of LLM vulnerabilities.

Harsh Chaudhari, Jamie Hayes, Matthew Jagielski et al. · deepmind

Model distillation has become essential for creating smaller, deployable language models that retain larger system capabilities. However, widespread deployment raises concerns about resilience to adversarial manipulation. This paper investigates vulnerability of distilled models to adversarial injection of biased content during training. We demonstrate that adversaries can inject subtle biases into teacher models through minimal data poisoning, which propagates to student models and becomes significantly amplified. We propose two propagation modes: Untargeted Propagation, where bias affects multiple tasks, and Targeted Propagation, focusing on specific tasks while maintaining normal behavior elsewhere. With only 25 poisoned samples (0.25% poisoning rate), student models generate biased responses 76.9% of the time in targeted scenarios - higher than 69.4% in teacher models. For untargeted propagation, adversarial bias appears 6x-29x more frequently in student models on unseen tasks. We validate findings across six bias types (targeted advertisements, phishing links, narrative manipulations, insecure coding practices), various distillation methods, and different modalities spanning text and code generation. Our evaluation reveals shortcomings in current defenses - perplexity filtering, bias detection systems, and LLM-based autorater frameworks - against these attacks. Results expose significant security vulnerabilities in distilled models, highlighting need for specialized safeguards. We propose practical design principles for building effective adversarial bias mitigation strategies.

Hanna Foerster, Ilia Shumailov, Yiren Zhao et al. · deepmind

Early research into data poisoning attacks against Large Language Models (LLMs) demonstrated the ease with which backdoors could be injected. More recent LLMs add step-by-step reasoning, expanding the attack surface to include the intermediate chain-of-thought (CoT) and its inherent trait of decomposing problems into subproblems. Using these vectors for more stealthy poisoning, we introduce ``decomposed reasoning poison'', in which the attacker modifies only the reasoning path, leaving prompts and final answers clean, and splits the trigger across multiple, individually harmless components. Fascinatingly, while it remains possible to inject these decomposed poisons, reliably activating them to change final answers (rather than just the CoT) is surprisingly difficult. This difficulty arises because the models can often recover from backdoors that are activated within their thought processes. Ultimately, it appears that an emergent form of backdoor robustness is originating from the reasoning capabilities of these advanced LLMs, as well as from the architectural separation between reasoning and final answer generation.

Nicolas Küchler, Ivan Petrov, Conrad Grobler et al. · deepmind

For nearly a decade the academic community has investigated backdoors in neural networks, primarily focusing on classification tasks where adversaries manipulate the model prediction. While demonstrably malicious, the immediate real-world impact of such prediction-altering attacks has remained unclear. In this paper we introduce a novel and significantly more potent class of backdoors that builds upon recent advancements in architectural backdoors. We demonstrate how these backdoors can be specifically engineered to exploit batched inference, a common technique for hardware utilization, enabling large-scale user data manipulation and theft. By targeting the batching process, these architectural backdoors facilitate information leakage between concurrent user requests and allow attackers to fully control model responses directed at other users within the same batch. In other words, an attacker who can change the model architecture can set and steal model inputs and outputs of other users within the same batch. We show that such attacks are not only feasible but also alarmingly effective, can be readily injected into prevalent model architectures, and represent a truly malicious threat to user privacy and system integrity. Critically, to counteract this new class of vulnerabilities, we propose a deterministic mitigation strategy that provides formal guarantees against this new attack vector, unlike prior work that relied on Large Language Models to find the backdoors. Our mitigation strategy employs a novel Information Flow Control mechanism that analyzes the model graph and proves non-interference between different user inputs within the same batch. Using our mitigation strategy we perform a large scale analysis of models hosted through Hugging Face and find over 200 models that introduce (unintended) information leakage between batch entries due to the use of dynamic quantization.

Ilia Shumailov, Daniel Ramage, Sarah Meiklejohn et al. · deepmind

We often interact with untrusted parties. Prioritization of privacy can limit the effectiveness of these interactions, as achieving certain goals necessitates sharing private data. Traditionally, addressing this challenge has involved either seeking trusted intermediaries or constructing cryptographic protocols that restrict how much data is revealed, such as multi-party computations or zero-knowledge proofs. While significant advances have been made in scaling cryptographic approaches, they remain limited in terms of the size and complexity of applications they can be used for. In this paper, we argue that capable machine learning models can fulfill the role of a trusted third party, thus enabling secure computations for applications that were previously infeasible. In particular, we describe Trusted Capable Model Environments (TCMEs) as an alternative approach for scaling secure computation, where capable machine learning model(s) interact under input/output constraints, with explicit information flow control and explicit statelessness. This approach aims to achieve a balance between privacy and computational efficiency, enabling private inference where classical cryptographic solutions are currently infeasible. We describe a number of use cases that are enabled by TCME, and show that even some simple classic cryptographic problems can already be solved with TCME. Finally, we outline current limitations and discuss the path forward in implementing them.

Luoxi Meng, Henry Feng, Ilia Shumailov et al.

Browser-using agents (BUAs) are an emerging class of AI agents that interact with web browsers in human-like ways, including clicking, scrolling, filling forms, and navigating across pages. While these agents help automate repetitive online tasks, they are vulnerable to prompt injection attacks that trick an agent into performing undesired actions, such as leaking private information or issuing unintended state-changing requests. We propose ceLLMate, a browser-level sandboxing framework that restricts the agent's ambient authority and reduces the blast radius of prompt injections. We address the semantic gap challenge that is fundamental to BUAs -- writing and enforcing security policies for low-level UI tools like clicks and keystrokes is brittle and error-prone. Our core insight is to perform sandboxing at the HTTP layer because all side-effecting UI operations will result in network communication to the website's backend. We implement ceLLMate as an agent-agnostic browser extension and demonstrate how it enables sandboxing policies that block prompt injection attacks in the WASP benchmark with 7.25--15% latency overhead.

Nils Philipp Walter, Chawin Sitawarin, Jamie Hayes et al.

Large Language Models (LLMs) are increasingly deployed in agentic systems that interact with an external environment; this makes them susceptible to prompt injections when dealing with untrusted data. To overcome this limitation, we propose SIC (Soft Instruction Control)-a simple yet effective iterative prompt sanitization loop designed for tool-augmented LLM agents. Our method repeatedly inspects incoming data for instructions that could compromise agent behavior. If such content is found, the malicious content is rewritten, masked, or removed, and the result is re-evaluated. The process continues until the input is clean or a maximum iteration limit is reached; if imperative instruction-like content remains, the agent halts to ensure security. By allowing multiple passes, our approach acknowledges that individual rewrites may fail but enables the system to catch and correct missed injections in later steps. Although immediately useful, worst-case analysis shows that SIC is not infallible; strong adversary can still get a 15% ASR by embedding non-imperative workflows. This nonetheless raises the bar.

Federico Barbero, Xiangming Gu, Christopher A. Choquette-Choo et al.

In this work, we show that it is possible to extract significant amounts of alignment training data from a post-trained model -- useful to steer the model to improve certain capabilities such as long-context reasoning, safety, instruction following, and maths. While the majority of related work on memorisation has focused on measuring success of training data extraction through string matching, we argue that embedding models are better suited for our specific goals. Distances measured through a high quality embedding model can identify semantic similarities between strings that a different metric such as edit distance will struggle to capture. In fact, in our investigation, approximate string matching would have severely undercounted (by a conservative estimate of $10\times$) the amount of data that can be extracted due to trivial artifacts that deflate the metric. Interestingly, we find that models readily regurgitate training data that was used in post-training phases such as SFT or RL. We show that this data can be then used to train a base model, recovering a meaningful amount of the original performance. We believe our work exposes a possibly overlooked risk towards extracting alignment data. Finally, our work opens up an interesting discussion on the downstream effects of distillation practices: since models seem to be regurgitating aspects of their training set, distillation can therefore be thought of as indirectly training on the model's original dataset.

Sven Gowal, Rudy Bunel, Florian Stimberg et al.

We introduce SynthID-Image, a deep learning-based system for invisibly watermarking AI-generated imagery. This paper documents the technical desiderata, threat models, and practical challenges of deploying such a system at internet scale, addressing key requirements of effectiveness, fidelity, robustness, and security. SynthID-Image has been used to watermark over ten billion images and video frames across Google's services and its corresponding verification service is available to trusted testers. For completeness, we present an experimental evaluation of an external model variant, SynthID-O, which is available through partnerships. We benchmark SynthID-O against other post-hoc watermarking methods from the literature, demonstrating state-of-the-art performance in both visual quality and robustness to common image perturbations. While this work centers on visual media, the conclusions on deployment, constraints, and threat modeling generalize to other modalities, including audio. This paper provides a comprehensive documentation for the large-scale deployment of deep learning-based media provenance systems.