OpenAI Launches Aardvark, an Agentic AI Security Researcher

OpenAI has introduced Aardvark, an AI-powered security researcher that autonomously finds, validates, and helps fix software vulnerabilities at scale. The system is currently in private beta, and OpenAI is accepting sign-ups for early testing.

Aardvark joins a small but growing field of agentic AI tools aimed at offensive and defensive security work. Earlier systems in this category—Google's Big Sleep and Project Zero's OSS-Fuzz automation among them—have shown that language models can surface real bugs in production code. OpenAI's entry signals that the company sees vulnerability research as a natural domain for agentic models, where multi-step reasoning, tool use, and autonomous verification matter more than single-shot question answering.

The three-stage scope is the notable design choice. Aardvark does not stop at detection, which is the stage where automated fuzzing and static analysis tools already operate. It also validates findings—filtering false positives that plague conventional scanners—and then moves to remediation, helping developers fix what it finds. Chaining those steps end to end is what distinguishes an agentic researcher from a traditional bug-finding tool.

The stakes are considerable. Thousands of critical vulnerabilities are disclosed every year, and security teams consistently report more open flaws than they can triage. If an autonomous system can compress the cycle from discovery to patch, it changes the economics of both defense and, potentially, exploitation—a dual-use tension that will shape how OpenAI gates access during the beta.

Details on model architecture, pricing, and availability timelines remain undisclosed. The private beta suggests a controlled rollout, consistent with how OpenAI has handled capability releases with security implications in the past. Researchers and security teams can request access through OpenAI's sign-up process.

Watch how OpenAI handles disclosure of vulnerabilities Aardvark discovers in the wild; that policy, more than the launch itself, will determine whether the tool strengthens the ecosystem's defenses or becomes a contested capability.