Google's Big Sleep agent found an exploitable SQLite bug that fuzzing missed
Google Project Zero and DeepMind built Big Sleep, an LLM agent for vulnerability research. It found a stack buffer underflow in a development version of SQLite before the code was released.
- Organization
- Google (Project Zero & DeepMind)
- Industry
- Security Research
- When
- Nov 2024
- Evidence
- Research paper
- Agents and tools
- Big SleepGemini 1.5 Pro
Reported outcome
Found a previously unknown, exploitable memory-safety bug in SQLite; developers patched it the same day, before any release.
- 0 (unreleased code)Users exposed
- Same dayTime to patch
Caveats
Described in a paper with a stated method. Peer review status varies.
- Big Sleep was a research project at the time, not a product.
- Google says traditional fuzzing did not find this bug, but one case does not show the agent beats fuzzing in general.
What the agent does
Researchers gave the agent a recently fixed SQLite bug as a starting point. The agent reviewed recent commits for a similar issue. It has a code browser, a debugger, a reporter tool, and a sandbox for Python scripts.
Why this works
This is variant analysis: "we found one bug, are there more like it?". It gives the agent a narrow, concrete target. The debugger lets it prove a crash instead of guessing.
What to copy
- Start from a known bug and ask for variants. It is much easier than "find any bug".
- Give the agent tools to prove findings (debugger, test harness), not only to read code.
- Run it on recent commits, so bugs are caught before release.
Starter prompt
Written by us. Use only on code you own or are authorized to test.
Context: commit {fix_commit} fixed this bug:
{bug_description}
Task: review commits since {date} in {repo} for code with a similar
root cause (same function family, same input handling pattern).
For each candidate:
1. Explain the root cause in 2 sentences.
2. Write a minimal test input that triggers it.
3. Run it under the sanitizer build and paste the output.
Report only candidates you reproduced.
Sources
- From Naptime to Big Sleep — Google Project Zero, Nov 2024
- Google says its AI found SQLite vulnerability that fuzzing missed — SecurityWeek, Nov 2024
Added Oct 16, 2026. Spotted an error? Send a correction.
Get new use cases weekly
One short email. Sources included.
Related use cases
Google DeepMind's CodeMender agent upstreamed 72 security fixes to open-source projects
72 security fixes upstreamed to open-source projects in six months, including projects as large as 4.5 million lines of code.
XBOW's autonomous pentesting agent reached #1 on HackerOne's US leaderboard
#1 on the HackerOne US leaderboard, with 1,000+ reports in a few months and 130+ confirmed and fixed by program owners.
Aider, the open-source coding agent, writes most of its own code
Aider's release notes report that Aider wrote up to 88% of the code in a release (v0.86.0, August 2025).