OpenAI Quadruples Bug Bounties, Hires Cathedral to Hack Its AI

What OpenAI is actually changing

Two things, and they point in the same direction: more eyes from outside, and more money to attract them.

The first is a partnership with Cathedral Systems, a security firm whose researchers will work alongside OpenAI's own team to test the company's AI infrastructure, find vulnerabilities and tighten defenses. The idea is fresh perspective — people who did not build the system looking at it the way an attacker would.

The second is the bug bounty. OpenAI did not publish the new dollar figures, but it says it is prepared to quadruple payouts for critical findings. The original program covered a wide surface: software patches, model privacy, threat detection and response, security tool integration and AI agent security.

That last item is the one to watch. AI agents that can act on their own — browsing, calling tools, touching data — are a much bigger attack surface than a chat window, and the industry is only starting to price that risk. Recent incidents like the ChatGPT account hijack that needed just one shared link show how quickly a small flaw can become a real problem.

Why outside hackers matter more than another internal sprint

OpenAI has run internal security pushes before, including asking a quarter of its developers to set aside time to hunt and fix weaknesses, using its own Codex tool to speed the work.

That approach has a blind spot: the people who built the system also built its assumptions. An outside firm like Cathedral is paid to disagree with those assumptions, and a bounty program extends that to anyone in the world with the skills and the patience.

The "traditional cyber defenses alone are no longer sufficient" line is the most interesting sentence in the announcement. It is the company saying out loud that the threat model has shifted — that AI systems can be probed, manipulated and attacked in ways firewalls and patch cycles were never designed to catch.

Cathedral CEO Gavin Kliger framed the partnership as an example for the rest of the industry, and he has a point. Anthropic has run a similar effort with its Project Glasswing, and using frontier models to test your own defenses is becoming the default posture rather than a marketing line.

The catch: no numbers, no timeline

A quadruple payout sounds generous until you know the baseline, and OpenAI did not share it.

Bug bounty programs live and die on their payout tables. Top-tier programs at major cloud and hardware companies pay six figures for the worst findings; smaller programs pay hundreds. Without the new figures, it is impossible to say whether OpenAI is now competitive with the best in the business or simply more competitive than it was last month.

The same goes for scope and timing. The announcement did not say when the higher rewards take effect, how Cathedral's findings will be disclosed, or whether the partnership is a one-off audit or an ongoing arrangement.

None of that makes the move meaningless. It just means the proof will arrive later, in the form of disclosed vulnerabilities and researchers deciding the payouts are worth their weekend.

What it means for you

If you use ChatGPT or build on OpenAI's API, the direct benefit is defensive: more independent testing should mean fewer exploitable flaws reaching production. There is nothing to sign up for and no price change attached to this announcement.

For security researchers in the US, UK and EU, the practical question is the new payout table. Until OpenAI publishes it, check the program's official page before spending time on a report — and note that bounty payments from a US company are taxable income in most jurisdictions.

OpenAI has made the right noises: outside testers, bigger rewards, and an honest admission that the old playbook is not enough. What is missing is the paperwork — the payout table, the scope and the disclosure terms. Those details will decide whether this is a serious security shift or a well-timed press release.