Skip to main content
Back

NUS Builds VeriGrey to Break AI Agents With 33% More Attack Success

A greybox fuzzing tool uses tool invocation sequences and context bridging to break through LLM agent defenses, finding 33% more injection attacks than black-box baselines.

Pro

  • VeriGrey applies greybox fuzzing to LLM agents, using tool invocation sequences as coverage feedback for search.
  • Its context-bridging mutator rewrites injection prompts to look like necessary steps of the user's benign task.
  • Beats black-box baseline by 33 percentage points on the AgentDojo benchmark using GPT-4.1.
  • Compromises Gemini CLI on 9 of 10 attacks including SSH key theft and cron persistence.
  • Triggers malicious OpenClaw skills 100% on Kimi-K2.5, 90% on Opus 4.6, 80% on GPT-5.2.
  • Tool-filter defenses reduce but do not eliminate attacks; VeriGrey still hits 17% under whitelisting.

Traditional fuzzers throw random inputs at software and watch for crashes. How do you fuzz an agent that reads natural language, calls APIs, and refuses anything that looks suspicious? A team from the National University of Singapore and the Max Planck Institute for Security and Privacy has built VeriGrey, a greybox fuzzing framework that treats the sequence of tool invocations as coverage feedback and uses LLM-crafted prompt mutations to smuggle malicious tasks past hardened defenses.

On the AgentDojo benchmark, VeriGrey achieves 33% additional efficacy in finding indirect prompt injection vulnerabilities with a GPT-4.1 back-end compared to a black-box baseline. In real-world case studies, it also surfaces prompts inducing several attack scenarios that black-box approaches miss on the widely used Gemini CLI coding agent and OpenClaw personal assistant.

Why fuzzing agents breaks the classic recipe

The classic greybox fuzzer, exemplified by AFL and libFuzzer, mutates byte-level inputs and uses branch coverage as a signal that a new input hit something interesting. That recipe collapses for agents. In conventional programs, branch coverage is reasonable feedback because each branch represents a distinct program behavior, but this correspondence weakens substantially for LLM agents whose actions are not captured by the branches exercised. A Gemini CLI run that calls read_file and one that calls write_file may execute the same wrapper code while doing wildly different things in the real world.

Prompts pose the second problem. Byte-level bit flips produce garbage that the LLM either ignores or refuses. Modern models have also been post-trained to spot obvious injections. If an injected task looks irrelevant to the user's original intent, the LLM can recognize and discard it. A fuzzer needs mutations that preserve fluency and slip past alignment.