Researchers used Claude to hack OpenAI

Same bounty, sharper stakes: researchers used an Anthropic cyber tool to reach OpenAI internal software information and suggest code changes through an employee account.

OpenAI paid $6,500, fixed the issues, and thanked the researchers. The timing—weeks after OpenAI agents hacked Hugging Face—keeps security and recursive improvement in the same headline frame.

Anthropic’s parallel disclosure that Claude leads 26% of R&D (from 1% in March) makes the recursive self-improvement debate less theoretical while a rival’s model is proving it can crack lab defenses.

We’re pairing the break-in with Anthropic’s R&D share number because they belong in one conversation.

FT/Ars emphasize what the employee account unlocked: private software information and suggested changes, with Codex wired to OpenAI’s GitHub organization. The entry remained Discourse on the community forum. Anthropic declined comment; OpenAI confirmed the fix.

Separately, Anthropic published that 26% of research and development work is now “led by” Claude—majority of tasks under human instruction and supervision—up from 1% in March. On 90% of tasks, AI still collaborates with a human. The company framed the data as helping the public see how close recursive self-improvement is. Models aren’t fully autonomous on the studied research yet. They are already doing most of the work on a rising share of tasks.

Useful AI that accelerates lab research is genuine progress. The harm if oversight lags: the same capability class that shortens exploit development also shortens the loop from model to next model. Temporary U.S. blocks on some Anthropic tools already showed how nervous governments are about cyber-capable releases. I’m watching whether recursive-R&D metrics and bug-bounty exploits force harder access gates—or just more press cycles.

Ars Technica / Financial Times, September 18, 2026. Complements TechCrunch’s Discourse/libheif technical path with FT’s GitHub-access emphasis and Anthropic’s R&D leadership stats.