Security Researchers Use Claude to Hack OpenAI in 72 Hours

A quick hack with AI in the middle

Security researchers used Claude, Anthropic's model, to hack into OpenAI in less than 72 hours. The team of three independent hackers at Hacktron managed to access employee accounts and the Monorepo repository, which reportedly contains "OpenAI's algorithmic secrets." The finding, reported by The Wall Street Journal, exposes a worrying fragility in the security of one of the most powerful AI companies in the world. The irony? They used one AI to breach another.

Context: when security gets outsourced

OpenAI, like many tech companies, outsources parts of its community infrastructure. Discourse, the service hosting its community forum, was the entry point. Researchers exploited this backdoor to escalate privileges and reach employee accounts. It's not the first time a third party has compromised a giant: in 2023, a similar attack hit GitHub and Okta. The difference now is speed and assistance from an advanced language model like Claude Opus 4.8 and 5.

Details: what they did and how

According to the WSJ, the Hacktron team:

• Exploited a vulnerability in Discourse to gain initial access.
• Used Claude Opus to automate the search for credentials and escalation paths.
• Accessed OpenAI's Monorepo on GitHub, though they didn't extract internal code.
• Sent a pull request from an employee's Codex account as proof of concept.

The whole process took less than three days. The researchers didn't damage systems or leak data, but they proved that the barrier between an external attacker and OpenAI's secrets is thinner than it seems.

Analysis: AI as a double-edged sword

This incident highlights a structural problem: AI companies are so obsessed with model security that they neglect employee and auxiliary service security. Anthropic didn't design Claude to hack, but its reasoning and code-generation capabilities make it a dangerous assistant in malicious hands. The winner here is Hacktron, gaining notoriety. The losers: OpenAI, whose security reputation is questioned, and to a lesser extent Anthropic, whose tool gets flagged.

My take: it's not a Claude failure, it's a human failure. AI accelerates what was already broken.

Implications: the future of cybersecurity with AI

OpenAI will likely review its third-party integrations and strengthen employee authentication. But the reputational damage is done. Beyond that, this case anticipates a new era of AI-assisted hacks, where execution speed multiplies. Companies will need to adopt AI-based defenses to counter. The arms race has begun, and this time both sides use language models.

The real risk isn't that OpenAI gets hacked, but that anyone with access to an LLM can do it.

Source: The Verge