Project Zero
Follow
From Naptime to Big Sleep: Using Large Language Models To Catch Vulnerabilities In Real-World Code
The Big Sleep team has discovered a previously unknown exploitable stack buffer underflow in SQLite, a widely used open-source database engine. The vulnerability was found using a collaboration between Google Project Zero and Google DeepMind, and was reported to the developers in early October. The issue was caused by a special sentinel value -1 used in an index-typed field, which led to a potential exploitable condition in a release build. The vulnerability was discovered by the Big Sleep agent, which uses a framework for large-language-model-assisted vulnerability research. The agent was able to find the vulnerability by analyzing a seed commit and applying pre-existing knowledge about SQLite. The Big Sleep team believes that this work has tremendous defensive potential, as it can help find vulnerabilities in software before they are even released.