Introducing Aardvark: OpenAI's Agentic Security Researcher

OpenAI has unveiled Aardvark (now renamed Codex Security), an agentic security researcher built on GPT-5. This agent helps developers and security teams detect and fix vulnerabilities in software codebases at scale.

Instead of relying on traditional fuzzing or software composition analysis (SCA) methods, Aardvark uses LLM-based reasoning and tool-use capabilities to read and analyze code like a human security researcher. It operates through the following multi-stage pipeline:

• Analysis: Analyzes the entire repository to generate a threat model reflecting the project's security goals and design.
• Commit scanning: Inspects changes whenever new code is committed, identifying vulnerabilities and explaining them step by step.
• Validation: Directly executes vulnerabilities in an isolated sandbox environment to confirm real-world exploitability and reduce false positives.
• Patching: Integrates with OpenAI Codex to generate fixes for discovered vulnerabilities, allowing users to apply patches with a single click.

Aardvark has demonstrated tangible results through internal testing and external partnerships. In benchmark tests, it showed a high recall rate, identifying 92% of known and synthetic vulnerabilities.

It has also been applied to open-source projects, resulting in the issuance of multiple CVEs. Going forward, OpenAI plans to contribute to ecosystem security by offering free scanning services for non-commercial open-source repositories.