Projects & tools

Featured Research

Selected projects from the lab, each with its own page covering the problem, our approach, and the papers, code, talks, and tools that came out of it. For the full list of our papers, see Publications.

ICLR 2026 Rank A*Black Hat USA 2026

GAVEL

Rule-based safety for LLMs — bringing signature-based cyber defense to AI.

GAVEL brings rule-based, signature-style cyber defense to AI safety: robust, auditable rules that monitor an LLM's internal activations and block unsafe behavior without retraining the model. ICLR 2026 (Rank A*), presented at Black Hat USA 2026.

AI SafetyAI Security
Explore the project
Featured in WIRED
USENIX Security 2026 Rank A* Distinguished Paper Award Runner-up

Love, Lies, and Language Models

How AI is entering romance-baiting scams, and why today's safeguards miss it.

Interviews with 145 scam-compound insiders, a blinded week-long conversation study, and a safeguard evaluation show that LLMs are already used in romance-baiting scams, out-perform human operators at building trust, and evade today's safety filters. USENIX Security 2026, Distinguished Paper Award runner-up.

Attacks Using AIAI Safety
Explore the project
EMNLP 2026 Rank A*

PRISM

Reading the instructions that steer an LLM, straight from its activations.

PRISM decodes a frozen LLM's hidden states into a faithful list of the instructions, constraints, and hidden objectives steering it, including prompt injections that never surface in the model's output. EMNLP 2026.

AI SafetyAI Security
Explore the project
ACSAC 2026 Rank A

Know Your Agent

Reconnaissance-driven pentesting of AI agents.

KYA brings the reconnaissance loop of real penetration testing to AI agents: it probes a black-box agent, extracts knowledge about its tools, task, and defenses, and uses that profile to craft far stronger indirect prompt injections. ACSAC 2026.

AI Security
Explore the project
NDSS 2026 Rank A*

Trust Me, I Know This Function

Hijacking LLM static analysis with Familiar Pattern Attacks.

Familiar Pattern Attacks exploit an abstraction bias in LLMs: tiny, behavior-preserving edits to familiar code make models misread what a program does. They transfer across models and languages and fool commercial code agents. NDSS 2026.

AI Security
Explore the project
USENIX Security 2025 Rank A*

Cloak, Honey, Trap

Proactive defenses that turn an AI hacking agent's weaknesses against it.

Cloak, Honey, Trap defends networks against autonomous LLM hacking agents by exploiting their biases, memory limits, and tokenization quirks: cloaking assets, exposing agents with LLM-specific honeytokens, and trapping them in loops or even making them run the defender's code. USENIX Security 2025.

Attacks Using AIAI Security
Explore the project