OpenAI launches Aardvark, an autonomous security research agent

OpenAI introduced Aardvark, an agent built on GPT-5 that continuously monitors a codebase’s commits and incoming changes, builds a threat model of it, and works to identify, explain and prioritise vulnerabilities much as a human security researcher would — rather than relying on the fuzzing or static-analysis techniques conventional scanning tools use. Where it finds an exploitable flaw, it uses OpenAI’s Codex coding tool to draft a patch for a human to review rather than auto-merging fixes itself.

OpenAI said Aardvark had been running against its own internal codebases and a set of external alpha partners, and credited it with surfacing at least ten vulnerabilities serious enough to be assigned official CVE identifiers in open-source projects. The company released the tool in private beta rather than general availability, saying it wanted to validate detection accuracy and reporting workflow with a limited set of partners before wider rollout.

The launch positioned Aardvark as a “defender-first” counterpart to the same large-model capabilities that have raised concern about AI-assisted offensive hacking, and drew comparisons to Google’s similarly aimed CodeMender tool announced weeks earlier. Both reflected a broader move by frontier labs to apply agentic coding models to security work specifically, on the argument that if language models can find exploitable bugs, the same capability should be turned toward finding and fixing them before attackers do — a case that remained to be tested at a scale beyond curated pilot programmes.