AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

OpenAI's Aardvark: AI Agent Finds Bugs in World's Most Audited Code

· AI + a16z

Matt Knight, OpenAI's VP of Security Products and Research (and former CISO for five years), joins a16z's Joel De La Garza to discuss Aardvark, an agentic AI security researcher that finds zero-day vulnerabilities and generates patches in the same exploratory way a human researcher would. Knight traces the arc from GPT-3, which couldn't analyze security logs or spot bugs in code, through GPT-4's breakthrough capabilities, to today's reasoning-enabled models. Aardvark hooks into a codebase, generates a threat model, explores for vulnerabilities, validates them in a sandbox, and produces ready-to-merge patches via Codex—already finding memory corruption bugs in highly audited targets like OpenSSH (CVE issued). Knight frames the work as creating defensive advantage at a time when open-source maintainers are vastly outgunned by nation-state attackers, referencing the XZ utils backdoor as a near-miss that could have compromised much of the internet. He argues AI scales security expertise to under-resourced developers rather than replacing the chronically short-staffed blue team workforce. Aardvark is in private beta with an open call to open-source maintainers.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.