Joshua Saxe

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri et al. · allen-ai, berkeley

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

Shengye Wan, Cyrus Nikolaidis, Daniel Song et al.

We are releasing a new suite of security benchmarks for LLMs, CYBERSECEVAL 3, to continue the conversation on empirically measuring LLM cybersecurity risks and capabilities. CYBERSECEVAL 3 assesses 8 different risks across two broad categories: risk to third parties, and risk to application developers and end users. Compared to previous work, we add new areas focused on offensive security capabilities: automated social engineering, scaling manual offensive cyber operations, and autonomous offensive cyber operations. In this paper we discuss applying these benchmarks to the Llama 3 models and a suite of contemporaneous state-of-the-art LLMs, enabling us to contextualize risks both with and without mitigations in place.

Manish Bhatt, Sahana Chennabasappa, Yue Li et al.

Large language models (LLMs) introduce new security risks, but there are few comprehensive evaluation suites to measure and reduce these risks. We present BenchmarkName, a novel benchmark to quantify LLM security risks and capabilities. We introduce two new areas for testing: prompt injection and code interpreter abuse. We evaluated multiple state-of-the-art (SOTA) LLMs, including GPT-4, Mistral, Meta Llama 3 70B-Instruct, and Code Llama. Our results show that conditioning away risk of attack remains an unsolved problem; for example, all tested models showed between 26% and 41% successful prompt injection tests. We further introduce the safety-utility tradeoff: conditioning an LLM to reject unsafe prompts can cause the LLM to falsely reject answering benign prompts, which lowers utility. We propose quantifying this tradeoff using False Refusal Rate (FRR). As an illustration, we introduce a novel test set to quantify FRR for cyberattack helpfulness risk. We find many LLMs able to successfully comply with "borderline" benign requests while still rejecting most unsafe requests. Finally, we quantify the utility of LLMs for automating a core cybersecurity task, that of exploiting software vulnerabilities. This is important because the offensive capabilities of LLMs are of intense interest; we quantify this by creating novel test sets for four representative problems. We find that models with coding capabilities perform better than those without, but that further work is needed for LLMs to become proficient at exploit generation. Our code is open source and can be used to evaluate other LLMs.

Sahana Chennabasappa, Cyrus Nikolaidis, Daniel Song et al.

Large language models (LLMs) have evolved from simple chatbots into autonomous agents capable of performing complex tasks such as editing production code, orchestrating workflows, and taking higher-stakes actions based on untrusted inputs like webpages and emails. These capabilities introduce new security risks that existing security measures, such as model fine-tuning or chatbot-focused guardrails, do not fully address. Given the higher stakes and the absence of deterministic solutions to mitigate these risks, there is a critical need for a real-time guardrail monitor to serve as a final layer of defense, and support system level, use case specific safety policy definition and enforcement. We introduce LlamaFirewall, an open-source security focused guardrail framework designed to serve as a final layer of defense against security risks associated with AI Agents. Our framework mitigates risks such as prompt injection, agent misalignment, and insecure code risks through three powerful guardrails: PromptGuard 2, a universal jailbreak detector that demonstrates clear state of the art performance; Agent Alignment Checks, a chain-of-thought auditor that inspects agent reasoning for prompt injection and goal misalignment, which, while still experimental, shows stronger efficacy at preventing indirect injections in general scenarios than previously proposed approaches; and CodeShield, an online static analysis engine that is both fast and extensible, aimed at preventing the generation of insecure or dangerous code by coding agents. Additionally, we include easy-to-use customizable scanners that make it possible for any developer who can write a regular expression or an LLM prompt to quickly update an agent's security guardrails.

Manish Bhatt, Sahana Chennabasappa, Cyrus Nikolaidis et al.

This paper presents CyberSecEval, a comprehensive benchmark developed to help bolster the cybersecurity of Large Language Models (LLMs) employed as coding assistants. As what we believe to be the most extensive unified cybersecurity safety benchmark to date, CyberSecEval provides a thorough evaluation of LLMs in two crucial security domains: their propensity to generate insecure code and their level of compliance when asked to assist in cyberattacks. Through a case study involving seven models from the Llama 2, Code Llama, and OpenAI GPT large language model families, CyberSecEval effectively pinpointed key cybersecurity risks. More importantly, it offered practical insights for refining these models. A significant observation from the study was the tendency of more advanced models to suggest insecure code, highlighting the critical need for integrating security considerations in the development of sophisticated LLMs. CyberSecEval, with its automated test case generation and evaluation pipeline covers a broad scope and equips LLM designers and researchers with a tool to broadly measure and enhance the cybersecurity safety properties of LLMs, contributing to the development of more secure AI systems.

Lauren Deason, Adam Bali, Ciprian Bejean et al.

Today's cyber defenders are overwhelmed by a deluge of security alerts, threat intelligence signals, and shifting business context, creating an urgent need for AI systems to enhance operational security work. While Large Language Models (LLMs) have the potential to automate and scale Security Operations Center (SOC) operations, existing evaluations do not fully assess the scenarios most relevant to real-world defenders. This lack of informed evaluation impacts both AI developers and those applying LLMs to SOC automation. Without clear insight into LLM performance in real-world security scenarios, developers lack a north star for development, and users cannot reliably select the most effective models. Meanwhile, malicious actors are using AI to scale cyber attacks, highlighting the need for open source benchmarks to drive adoption and community-driven improvement among defenders and model developers. To address this, we introduce CyberSOCEval, a new suite of open source benchmarks within CyberSecEval 4. CyberSOCEval includes benchmarks tailored to evaluate LLMs in two tasks: Malware Analysis and Threat Intelligence Reasoning--core defensive domains with inadequate coverage in current benchmarks. Our evaluations show that larger, more modern LLMs tend to perform better, confirming the training scaling laws paradigm. We also find that reasoning models leveraging test time scaling do not achieve the same boost as in coding and math, suggesting these models have not been trained to reason about cybersecurity analysis, and pointing to a key opportunity for improvement. Finally, current LLMs are far from saturating our evaluations, showing that CyberSOCEval presents a significant challenge for AI developers to improve cyber defense capabilities.

Tomek Korbak, Mikita Balesni, Elizabeth Barnes et al. · deepmind

AI systems that "think" in human language offer a unique opportunity for AI safety: we can monitor their chains of thought (CoT) for the intent to misbehave. Like all other known AI oversight methods, CoT monitoring is imperfect and allows some misbehavior to go unnoticed. Nevertheless, it shows promise and we recommend further research into CoT monitorability and investment in CoT monitoring alongside existing safety methods. Because CoT monitorability may be fragile, we recommend that frontier model developers consider the impact of development decisions on CoT monitorability.

Younghoo Lee, Joshua Saxe, Richard Harang

Targeted phishing emails are on the rise and facilitate the theft of billions of dollars from organizations a year. While malicious signals from attached files or malicious URLs in emails can be detected by conventional malware signatures or machine learning technologies, it is challenging to identify hand-crafted social engineering emails which don't contain any malicious code and don't share word choices with known attacks. To tackle this problem, we fine-tune a pre-trained BERT model by replacing the half of Transformer blocks with simple adapters to efficiently learn sophisticated representations of the syntax and semantics of the natural language. Our Context-Aware network also learns the context representations between email's content and context features from email headers. Our CatBERT(Context-Aware Tiny Bert) achieves a 87% detection rate as compared to DistilBERT, LSTM, and logistic regression baselines which achieve 83%, 79%, and 54% detection rates at false positive rates of 1%, respectively. Our model is also faster than competing transformer approaches and is resilient to adversarial attacks which deliberately replace keywords with typos or synonyms.

Ben Gelman, Bryan Hoyle, Jessica Moore et al.

Code search and comprehension have become more difficult in recent years due to the rapid expansion of available source code. Current tools lack a way to label arbitrary code at scale while maintaining up-to-date representations of new programming languages, libraries, and functionalities. Comprehensive labeling of source code enables users to search for documents of interest and obtain a high-level understanding of their contents. We use Stack Overflow code snippets and their tags to train a language-agnostic, deep convolutional neural network to automatically predict semantic labels for source code documents. On Stack Overflow code snippets, we demonstrate a mean area under ROC of 0.957 over a long-tailed list of 4,508 tags. We also manually validate the model outputs on a diverse set of unlabeled source code documents retrieved from Github, and we obtain a top-1 accuracy of 86.6%. This strongly indicates that the model successfully transfers its knowledge from Stack Overflow snippets to arbitrary source code documents.

Ethan M. Rudd, Richard Harang, Joshua Saxe

Malicious email attachments are a growing delivery vector for malware. While machine learning has been successfully applied to portable executable (PE) malware detection, we ask, can we extend similar approaches to detect malware across heterogeneous file types commonly found in email attachments? In this paper, we explore the feasibility of applying machine learning as a static countermeasure to detect several types of malicious email attachments including Microsoft Office documents and Zip archives. To this end, we collected a dataset of over 5 million malicious/benign Microsoft Office documents from VirusTotal for evaluation as well as a dataset of benign Microsoft Office documents from the Common Crawl corpus, which we use to provide more realistic estimates of thresholds for false positive rates on in-the-wild data. We also collected a dataset of approximately 500k malicious/benign Zip archives, which we scraped using the VirusTotal service, on which we performed a separate evaluation. We analyze predictive performance of several classifiers on each of the VirusTotal datasets using a 70/30 train/test split on first seen time, evaluating feature and classifier types that have been applied successfully in commercial antimalware products and R&D contexts. Using deep neural networks and gradient boosted decision trees, we are able to obtain ROC curves with > 0.99 AUC on both Microsoft Office document and Zip archive datasets. Discussion of deployment viability in various antimalware contexts is provided.

Joshua Saxe, Richard Harang, Cody Wild et al.

Malicious web content is a serious problem on the Internet today. In this paper we propose a deep learning approach to detecting malevolent web pages. While past work on web content detection has relied on syntactic parsing or on emulation of HTML and Javascript to extract features, our approach operates directly on a language-agnostic stream of tokens extracted directly from static HTML files with a simple regular expression. This makes it fast enough to operate in high-frequency data contexts like firewalls and web proxies, and allows it to avoid the attack surface exposure of complex parsing and emulation code. Unlike well-known approaches such as bag-of-words models, which ignore spatial information, our neural network examines content at hierarchical spatial scales, allowing our model to capture locality and yielding superior accuracy compared to bag-of-words baselines. Our proposed architecture achieves a 97.5% detection rate at a 0.1% false positive rate, and classifies small-batched web pages at a rate of over 100 per second on commodity hardware. The speed and accuracy of our approach makes it appropriate for deployment to endpoints, firewalls, and web proxies.

Joshua Saxe, Konstantin Berlin

For years security machine learning research has promised to obviate the need for signature based detection by automatically learning to detect indicators of attack. Unfortunately, this vision hasn't come to fruition: in fact, developing and maintaining today's security machine learning systems can require engineering resources that are comparable to that of signature-based detection systems, due in part to the need to develop and continuously tune the "features" these machine learning systems look at as attacks evolve. Deep learning, a subfield of machine learning, promises to change this by operating on raw input signals and automating the process of feature design and extraction. In this paper we propose the eXpose neural network, which uses a deep learning approach we have developed to take generic, raw short character strings as input (a common case for security inputs, which include artifacts like potentially malicious URLs, file paths, named pipes, named mutexes, and registry keys), and learns to simultaneously extract features and classify using character-level embeddings and convolutional neural network. In addition to completely automating the feature design and extraction process, eXpose outperforms manual feature extraction based baselines on all of the intrusion detection problems we tested it on, yielding a 5%-10% detection rate gain at 0.1% false positive rate compared to these baselines.

Konstantin Berlin, Joshua Saxe

Enterprise networks are in constant danger of being breached by cyber-attackers, but making the decision about what security tools to deploy to mitigate this risk requires carefully designed evaluation of security products. One of the most important metrics for a protection product is how well it is able to stop malware, specifically on "zero"-day malware that has not been seen by the security community before. However, evaluating zero-day performance is difficult, because of larger number of previously unseen samples that are needed to properly measure the true and false positive rate, and the challenges involved in accurately labeling these samples. This paper addresses these issues from a statistical and practical perspective. Our contributions include first showing that the number of benign files needed for proper evaluation is on the order of a millions, and the number of malware samples needed is on the order of tens of thousands. We then propose and justify a time-delay method for easily collecting large number of previously unseen, but labeled, samples. This enables cheap and accurate evaluation of zero-day true and false positive rates. Finally, we propose a more fine-grain labeling of the malware/benignware in order to better model the heterogeneous distribution of files on various networks.

Joshua Saxe, Rafael Turner, Kristina Blokhin

In this paper we introduce CrowdSource, a statistical natural language processing system designed to make rapid inferences about malware functionality based on printable character strings extracted from malware binaries. CrowdSource "learns" a mapping between low-level language and high-level software functionality by leveraging millions of web technical documents from StackExchange, a popular network of technical question and answer sites, using this mapping to infer malware capabilities. This paper describes our approach and provides an evaluation of its accuracy and performance, demonstrating that it can detect at least 14 high-level malware capabilities in unpacked malware binaries with an average per-capability f-score of 0.86 and at a rate of tens of thousands of binaries per day on commodity hardware.

Joshua Saxe, Konstantin Berlin

Malware remains a serious problem for corporations, government agencies, and individuals, as attackers continue to use it as a tool to effect frequent and costly network intrusions. Machine learning holds the promise of automating the work required to detect newly discovered malware families, and could potentially learn generalizations about malware and benign software that support the detection of entirely new, unknown malware families. Unfortunately, few proposed machine learning based malware detection methods have achieved the low false positive rates required to deliver deployable detectors. In this paper we a deep neural network malware classifier that achieves a usable detection rate at an extremely low false positive rate and scales to real world training example volumes on commodity hardware. Specifically, we show that our system achieves a 95% detection rate at 0.1% false positive rate (FPR), based on more than 400,000 software binaries sourced directly from our customers and internal malware databases. We achieve these results by directly learning on all binaries, without any filtering, unpacking, or manually separating binary files into categories. Further, we confirm our false positive rates directly on a live stream of files coming in from Invincea's deployed endpoint solution, provide an estimate of how many new binary files we expected to see a day on an enterprise network, and describe how that relates to the false positive rate and translates into an intuitive threat score. Our results demonstrate that it is now feasible to quickly train and deploy a low resource, highly accurate machine learning classification model, with false positive rates that approach traditional labor intensive signature based methods, while also detecting previously unseen malware.

Konstantin Berlin, David Slater, Joshua Saxe

As antivirus and network intrusion detection systems have increasingly proven insufficient to detect advanced threats, large security operations centers have moved to deploy endpoint-based sensors that provide deeper visibility into low-level events across their enterprises. Unfortunately, for many organizations in government and industry, the installation, maintenance, and resource requirements of these newer solutions pose barriers to adoption and are perceived as risks to organizations' missions. To mitigate this problem we investigated the utility of agentless detection of malicious endpoint behavior, using only the standard build-in Windows audit logging facility as our signal. We found that Windows audit logs, while emitting manageable sized data streams on the endpoints, provide enough information to allow robust detection of malicious behavior. Audit logs provide an effective, low-cost alternative to deploying additional expensive agent-based breach detection systems in many government and industrial settings, and can be used to detect, in our tests, 83% percent of malware samples with a 0.1% false positive rate. They can also supplement already existing host signature-based antivirus solutions, like Kaspersky, Symantec, and McAfee, detecting, in our testing environment, 78% of malware missed by those antivirus systems.