Big Sleep — the first agent-found memory-safety flaw in real software

The museum's security wing, until this object, was the attackers' wing: DAN, prompt injection, adversarial suffixes, the models that stole the answer key. This is the defenders' first entry — the same class of agent, pointed at the same code attackers read, arriving first.

Big Sleep did not out-fuzz the fuzzers. It read recent commits the way a human researcher reads them — asking what related mistake might still be alive nearby — and found an exploitable stack buffer underflow in SQLite that twenty years of fuzzing had not surfaced. Reported in early October 2024; fixed the same day; never shipped in a release. A vulnerability corrected before it exists in anyone's hands collapses the attacker's window to zero — the only patch cycle with no victims is the one that finishes before release.

The claim survives because its authors kept it narrow: first *public* example, *exploitable*, *memory-safety*, *widely used real-world software* — with Team Atlanta's earlier AIxCC find cited in the same breath, and their own results called highly experimental. A first stated with its boundary is the only kind this museum accessions.