OpenAI brings Hacktron, the team that broke in with Claude, inside to probe its third-party seams

OpenAI announced a partnership with Hacktron to hunt vulnerabilities in its own systems, with jointly selected high-priority test areas and a stated focus on how OpenAI systems interact with third-party services. Hacktron also gets expanded access to frontier models via Daybreak Red, part of OpenAI’s broader push to put frontier cyber models in defenders’ hands (Daybreak). OpenAI says it knew Hacktron through the bug bounty program and was impressed by the impact of its disclosures. One of them is well documented: Hacktron’s small team used Anthropic’s Claude to chain Discourse and libheif flaws into OpenAI employee ChatGPT and Codex access, earning a bounty (the Claude-assisted break-in).

Hacktron builds AI agents for code review, penetration testing and red teaming, and its public work targets the seams between apps, cloud infrastructure and identity systems. That is where agent-era risk concentrates: identity, cloud and application boundaries, not only model weights or prompt-injection demos. CEO Zayne Zhang frames the bet as human creativity plus AI-driven offensive capability, with frontier cyber models in researchers’ hands, matching OpenAI’s stated plan to accelerate defense with frontier AI and apply that urgency inward. OpenAI’s own specialist model shows how capable these tools are getting (GPT-6 Cyber).

The same week OpenAI flagged Cathedral researchers joining its next security sprint and higher bug-bounty rewards for independent finders. Internal review alone cannot secure an interconnected platform moving at agent speed.

Labs inviting serious offensive partners and raising bounties is good: security AI should find bugs before attackers do. The risk is partnerships as press-cycle armor, with scoped tests that avoid the riskiest third-party paths or model access that outruns the fix pipeline. Announcements are easy; mean time to remediate after a Hacktron or Cathedral finding is the scoreboard, along with what lands in bounty scope and what gets patched. OpenAI’s recent run of agent-safety incidents raises the stakes (training pause). More in our OpenAI coverage and cybersecurity coverage.