OpenAI partners with Hacktron to test its third-party connections

Two known bugs in two public tools let a stranger open a pull request inside OpenAI's internal monorepo. That stranger was Hacktron, a security firm whose researchers chained the two flaws in July and then stopped.

OpenAI has made Hacktron a formal partner for vulnerability research. The work will focus on how OpenAI's systems connect to third-party services.

What Hacktron does

The company builds AI agents for code review, penetration testing and red teaming. Its researchers look for flaws where applications, cloud infrastructure and identity systems meet.

OpenAI engineers and Hacktron researchers will pick high-priority areas to test together under the partnership. OpenAI said it was impressed by the impact of the vulnerabilities Hacktron found during the Security Bug Bounty programme.

What the Daybreak tiers mean

Hacktron also gets access to a security-tier set of OpenAI models called Daybreak Red. Unite.ai writes that the tier offers purpose-trained cybersecurity models for authorised vulnerability research, exploit validation and security testing. The default Daybreak Blue tier offers general frontier models with defensive safeguards instead.

The demonstration that preceded the partnership

Hacktron had already proved what it could find. Earlier this year its researchers chained two critical flaws. One was a remote code execution bug in the Discourse forum's image processing. The other was a single-sign-on flaw at OpenAI.

Using that chain, the researchers accessed an OpenAI employee's Codex account. They then used Codex to open a harmless pull request in OpenAI's internal monorepo. After proving the access they stopped and reported both flaws.

OpenAI confirmed a fix about 14 hours after the report, per unite.ai. Discourse patched its side within two days. Hacktron received a 6,500 dollar bounty from OpenAI.

What the partnership guards against

Partnership scope includes joint testing, human oversight, scoped permissions, identity verification and monitoring. Isolated workflows stay apart from production systems.

The Codex pull request is the part worth studying for anyone building an agent that holds a developer's credentials. An agent given tokens or single-sign-on access to a developer platform can do anything that account can do. The attack never touched the model. A chained flaw in a forum and a shared image library was enough to reach the code. Hacktron's own pitch is that both sides of security work are using AI agents now. This arrangement puts that pitch inside one of the frontier labs.

Source

Partnering with Hacktron to find vulnerabilities in OpenAI's systems, OpenAI. Partnering with OpenAI, Zayne Zhang, Hacktron. OpenAI Partners With Hacktron to Test Third-Party Service Connections, Miles Okada, Unite.ai.