Originally published on Rohan Tech Official.

OpenAI just hired a company to break into OpenAI. On November 2, the company announced that Cathedral Systems — an AI security startup founded by ex-DOGE staffers — will join its next internal security sprint, using OpenAI's own models to attack its defenses. At the same time, OpenAI more than quadrupled its bug bounties, with top payouts reaching $172,000.

Why would the world's most-watched AI lab invite hackers inside? Because the threats are getting smarter — and so are the defenses it needs.

Who Is Cathedral?

Cathedral Systems is a young AI security startup with an unusual pedigree: its founders are former staffers from DOGE, the government efficiency outfit. But the detail that turned heads came from the Washington Post, which reported that Cathedral's AI agents got inside Anthropic's internal systems in October — one of the most security-conscious AI labs on the planet.

That's quite a calling card. If your agents can breach Anthropic, every other AI company wants to know how you did it — ideally before someone with worse intentions figures it out. For more on OpenAI's recent moves, see our coverage of the ChatGPT Codex single shared plan.

Attacking OpenAI With Its Own Models

Here's the twist that makes this sprint genuinely interesting: Cathedral won't just throw human red-teamers at OpenAI's infrastructure. It will use OpenAI's own models to attack OpenAI's defenses. AI agents probing AI systems, hunting for weaknesses at machine speed.

This is where AI security is heading. Human testers are thorough but slow; AI agents can try thousands of attack paths in the time it takes a person to write one report. The same agentic capabilities that make AI useful — planning, tool use, persistence — make it a formidable attacker. OpenAI is essentially stress-testing its fortress with an army that thinks like the enemy.

Bug Bounties, Quadrupled

Alongside the Cathedral announcement, OpenAI massively raised the stakes for independent researchers. Critical vulnerabilities in ChatGPT, Codex, and the API now pay up to $30,000. Exceptional findings start at a $100,000 base — and with quality, loyalty, and campaign bonuses, top payouts can reach $172,000.

That's more than four times the previous rewards, and it puts OpenAI's program among the most generous in tech. The message to the security community is unmistakable: we want your best work, and we'll pay for it. Researchers can find the full details on OpenAI's Bugcrowd page or through OpenAI's official site.

Why Now?

Timing is everything. AI systems are becoming more capable, more autonomous, and more deeply embedded in critical workflows — which means the consequences of a breach keep growing. An attacker who compromises a frontier model's training pipeline or deployment infrastructure doesn't just steal data; they could poison the model itself.

The Cathedral-Anthropic episode in October likely focused minds. When a startup's agents can penetrate a top-tier lab, it proves the threat model has changed. Human red teams alone can't keep up with agentic attacks — you need agentic defenses, tested by agentic attackers. The Anthropic text watermark story showed how seriously labs take integrity; this sprint shows how seriously they take security.

What This Tells Us About AI Security

Three takeaways. First, the AI security market is maturing fast — startups like Cathedral can now sell "our agents broke into Anthropic" as a credential. Second, the line between AI capability research and AI security research is dissolving; the same agents that do useful work can do harmful work, and defending against them requires understanding both.

Third, money talks. A $172K top bounty will pull elite researchers toward AI targets and away from traditional software. Expect the next year to surface vulnerabilities in AI systems that nobody has thought to look for yet — because until now, nobody was paying enough to look.

What's Next

The sprint itself will run behind closed doors, and OpenAI will decide what to disclose. Watch for two things: whether the bounty program produces headline findings in the coming months, and whether other labs copy the playbook.

If Cathedral's agents find real holes in OpenAI's defenses, that's a win for everyone who uses these systems. The uncomfortable alternative — that nobody looked — is no longer an option.