Vulnerability research on open-source code
An agent studies code, writes test inputs and runs them in a sandbox to find exploitable bugs, which researchers confirm and report.
The use case
Finding memory-safety bugs in large codebases takes skilled researchers a long time. Google Project Zero and DeepMind's Big Sleep agent reads code, forms hypotheses about bugs, writes test inputs and runs them in a sandbox with debugging tools, similar to how a human researcher works. Google reported it found an exploitable bug in SQLite before release. Teams can use the same approach for variant analysis of known bug patterns. Researchers still reproduce each finding, judge exploitability, and handle disclosure through proper channels.
Example workflow
- Choose a target and a known bug pattern
- Agent reviews code and writes test inputs
- Run tests in an isolated sandbox
- Researcher reproduces and assesses findings
- Report through coordinated disclosure
Tools
- Big Sleep
Risks and guardrails
- Run everything in isolated sandboxes
- False positives waste maintainer time; confirm first
- Follow responsible disclosure
Sources
Written by Agents Lookup. Credits and official pages for the tools this use case names:
- Project Zero: from Naptime to Big Sleep: projectzero.google/2024/10/from-naptime-to-big-sleep.html