Google Big Sleep

Big Sleep examines code and tests suspected flaws

Big Sleep is an AI vulnerability research agent from Google DeepMind and Project Zero. It developed from Project Naptime, which connected a language model to tools for reading code, debugging and testing inputs. Google’s introduction describes using recent code changes as starting points for finding related flaws.

Its first announced real-world finding was a stack buffer underflow in SQLite. Google reported it in early October 2024, and SQLite fixed it the same day, before an official release. Project Zero published the account on 1 November 2024.

The register contains 26 attributed entries

The AI-found vulnerability register contains 26 Big Sleep entries as checked on 8 October 2026. Eighteen have credits in official CVE records, seven have credits in vendor advisories, and one is the original SQLite finding described by Google.

These are individually attributed entries. Read each linked record for its disclosure date, affected software, severity and finder credit. A programme-wide claim can include findings at a different stage of disclosure.

Human contributions remain part of the credit

The credit for CVE-2025-6965, an integer-truncation flaw in SQLite, names Vlad Stolyarov of Google’s Threat Analysis Group with assistance from Big Sleep. Google says threat intelligence and the agent helped identify the vulnerability before threat actors used it.

Other records credit Big Sleep directly. Those credits establish attribution; the amount of human involvement follows the evidence published for each finding. The register keeps the human-assisted SQLite result distinct from findings described as agent-led.

curl records a $505 award

The curl advisory for CVE-2025-9086 credits Google Big Sleep and records a $505 award. Published on 10 September 2025, it describes an out-of-bounds read in cookie-path handling, rates the issue low and identifies curl 8.16.0 as the fixed version.

The award is recorded against that specific finding in the register. Reward amounts, eligibility and research costs are separate measurements. The bounty economics tracker compares programmes’ advertised terms and maximum rewards.

Disclosure follows the published policy

Project Zero’s July 2025 policy update explicitly includes Big Sleep in its reporting-transparency trial. It retains the 90+30 policy: a 90-day period for the vendor to fix a reported bug, with a 30-day patch-adoption period when the fix arrives before the deadline.

The trial also publishes the affected vendor, product, report date and disclosure deadline approximately a week after reporting. The issue’s own timeline is the source for those dates. Downstream users need the fixed version in the product they actually run.