Julien Piet

Samuel Pfrommer, Brendon G. Anderson, Julien Piet et al.

Recent works have introduced input-convex neural networks (ICNNs) as learning models with advantageous training, inference, and generalization properties linked to their convex structure. In this paper, we propose a novel feature-convex neural network architecture as the composition of an ICNN with a Lipschitz feature map in order to achieve adversarial robustness. We consider the asymmetric binary classification setting with one "sensitive" class, and for this class we prove deterministic, closed-form, and easily-computable certified robust radii for arbitrary $\ell_p$-norms. We theoretically justify the use of these models by characterizing their decision region geometry, extending the universal approximation theorem for ICNN regression to the classification setting, and proving a lower bound on the probability that such models perfectly fit even unstructured uniformly distributed data in sufficiently high dimensions. Experiments on Malimg malware classification and subsets of MNIST, Fashion-MNIST, and CIFAR-10 datasets show that feature-convex classifiers attain state-of-the-art certified $\ell_1$-radii as well as substantial $\ell_2$- and $\ell_{\infty}$-radii while being far more computationally efficient than any competitive baseline.

Julien Piet, Maha Alrashed, Chawin Sitawarin et al. · pku

Large Language Models (LLMs) are attracting significant research attention due to their instruction-following abilities, allowing users and developers to leverage LLMs for a variety of tasks. However, LLMs are vulnerable to prompt-injection attacks: a class of attacks that hijack the model's instruction-following abilities, changing responses to prompts to undesired, possibly malicious ones. In this work, we introduce Jatmo, a method for generating task-specific models resilient to prompt-injection attacks. Jatmo leverages the fact that LLMs can only follow instructions once they have undergone instruction tuning. It harnesses a teacher instruction-tuned model to generate a task-specific dataset, which is then used to fine-tune a base model (i.e., a non-instruction-tuned model). Jatmo only needs a task prompt and a dataset of inputs for the task: it uses the teacher model to generate outputs. For situations with no pre-existing datasets, Jatmo can use a single example, or in some cases none at all, to produce a fully synthetic dataset. Our experiments on seven tasks show that Jatmo models provide similar quality of outputs on their specific task as standard LLMs, while being resilient to prompt injections. The best attacks succeeded in less than 0.5% of cases against our models, versus 87% success rate against GPT-3.5-Turbo. We release Jatmo at https://github.com/wagner-group/prompt-injection-defense.

Evan Luo, Julien Piet, David Wagner

Security analysts routinely query system logs to detect threats and investigate incidents, but each log source uses its own semi-structured format: logs are cheap to produce, but expensive to use. The standard approach, building per-source parsers to normalize logs into structured schemas, is powerful but requires continuous engineering effort for each new format. Querying raw logs directly with tools like grep avoids this cost, but requires analysts to know each source's message variants and cannot express the multi-line temporal queries that security investigations demand. We present Sieve, a system that generates executable query code from natural-language security questions by grounding a large language model with lightweight, automatically extracted log-format context, requiring only one LLM call per query followed by deterministic execution. Evaluating 133 security queries across 5 log types, we find that Sieve achieves over a 3x reduction in error rate on complex temporal and cross-event queries compared to manual analyst scripting, with the largest gains on the multi-line correlation tasks most critical to active investigations. Our results and benchmark provide evidence that LLM-generated code can bridge the gap between the expressiveness of structured log querying and the immediacy of working directly with raw files.

Julien Piet, Annabella Chow, Yiwei Hou et al.

ReAct has become the default architecture across LLM agents, and many existing web agents follow this paradigm. We argue that it is the wrong default for web agents. Instead, web agents should default to plan-then-execute: commit to a task-specific program before observing runtime web content, then execute it. The reason is that web content mixes inputs from many parties. An e-commerce product page may combine a seller's listing, customer reviews and sponsored advertisements. Under ReAct, all of this content flows into the model when deciding on the next action, creating a direct path for prompt injections to steer the agent's control flow. Plan-then-execute changes this boundary: untrusted data may influence values or branches inside a predefined execution graph, but it cannot redefine the user task or cause the model to synthesize new actions at runtime. We analyze WebArena, a popular web agent benchmark, and find that all tasks are compatible with plan-then-execute, while 80% can be completed with a purely programmatic plan, without any runtime LLM subroutine. We identify the main barrier to adopting plan-then-execute on the web: For it to work well, tools must map cleanly to semantic actions, with effects known before execution, so agents have enough information to plan. The web does not naturally expose that interface. Browser tools such as click, type, and scroll have page-dependent meanings. Planning at this layer is near-sighted: the agent can only see actions on the current page, and later actions appear only after it acts. Closing this gap requires typed interfaces that turn website interactions from clicks and keystrokes to task-level operations. This is an infrastructure problem, not a modeling problem. Web tasks do not need reactivity by default; they need typed, complete, auditable website APIs.

Debeshee Das, Julien Piet, Darya Kaviani et al.

Memory systems enable otherwise-stateless LLM agents to persist user information across sessions, but also introduce a new attack surface. We characterize the Trojan Hippo attack, a class of persistent memory attacks that operates in a more realistic threat model than prior memory poisoning work: the attacker plants a dormant payload into an agent's long-term memory via a single untrusted tool call (e.g., a crafted email), which activates only when the user later discusses sensitive topics such as finance, health, or identity, and exfiltrates high-value personal data to the attacker. While anecdotal demonstrations of such attacks have appeared against deployed systems, no prior work systematically evaluates them across heterogeneous memory architectures and defenses.We introduce a dynamic evaluation framework comprising two components: (1) an OpenEvolve-based adaptive red-teaming benchmark that stress-tests defenses and memory backends against continuously refined attacks, and (2) the first capability-aware security/utility analysis for persistent memory systems, enabling principled reasoning about defense deployment across different usage profiles. Instantiated on an email assistant across four memory backends (explicit tool memory, agentic memory, RAG, and sliding-window context), Trojan Hippo achieves up to 85-100 percent ASR against current frontier models from OpenAI and Google, with planted memories successfully activating even after 100 benign sessions. We evaluate four memory-system defenses inspired by basic security principles, finding they substantially reduce attack success rates (to as low as 0-5 percent), though at utility costs that vary widely with task requirements. Because of this substantial security-utility tradeoff, the effective real-world deployment of defenses remains an open challenge, which our evaluation framework is specifically designed to address.