OpenAI tests Aardvark, an AI security agent

OpenAI has introduced Aardvark, an artificial intelligence agent built on GPT-5 that analyzes code to find vulnerabilities, check whether they can be exploited, and propose fixes. For now, it is available only to a small group of partners in a private beta.

The tool is designed to work continuously alongside development teams. It reviews changes arriving in a repository, compares each modification with the rest of the project, and looks for flaws that could open the door to an attack.

It does not simply flag suspicious lines. Aardvark tries to understand how the software works, much like a human security researcher would: it reads the code, studies its design, writes tests, and uses tools to check its hypotheses.

How Aardvark works

The system goes through several stages before alerting the team:

• Analyzes the entire project to create a model of its security objectives and the most sensitive parts of its design.
• Reviews every code change and can also examine the repository's history when it connects for the first time.
• Explains the potential flaw step by step, with annotations that a person can review.
• Attempts to reproduce the attack in an isolated environment to check whether the vulnerability can really be exploited and reduce false positives.
• Proposes a patch through Codex, OpenAI's programming system. The change is attached to the report so a developer can review it and decide whether to apply it.

Unlike traditional tools such as fuzzing, which tests a program with large quantities of inputs, or third-party component analysis, Aardvark relies mainly on a language model's reasoning and its use of tools. The goal is to detect problems that depend on several conditions and can go unnoticed in simpler automated reviews.

What results it has achieved

OpenAI says Aardvark has been running continuously for several months in its own repositories and those of external partners. During that time, it has found vulnerabilities, logic errors, incomplete fixes, and privacy issues.

In tests using repositories prepared to measure its performance, it identified 92% of known and artificially created vulnerabilities. That result applies to those evaluation repositories, not as a guarantee for every project.

The tool has also been used on open-source projects. OpenAI says it discovered and responsibly disclosed several vulnerabilities, ten of them with CVE identifiers, the system used to catalog known security flaws.

Why it matters to you

Software supports banks, hospitals, government agencies, stores, and digital services. More than 40,000 CVEs were recorded in 2024, according to data cited by OpenAI. Its tests also indicate that around 1.2% of code changes introduce errors.

That means a flaw can enter an application through a small modification and remain hidden until someone exploits it. Detecting it while the code is still being developed is usually faster and cheaper than fixing it after an incident.

For teams, Aardvark does not automatically replace experts or apply changes without supervision. Its role is to monitor continuously, provide evidence about the risk, and prepare a possible solution within tools they already use, such as GitHub and Codex.

OpenAI plans to offer free analysis for certain noncommercial open-source repositories. The company has also updated its coordinated disclosure policy to give developers more time and encourage collaboration when a flaw appears.

The private beta will help determine how far detection goes in real-world projects, how much noise it generates, and whether the proposed patches are useful. The important thing is not for AI to find every error on its own, but to turn a one-off security review into continuous monitoring as software changes.