AI vs AI: Claude Hacked OpenAI So Fast, Even Vitalik Said "LMAO

Independent researchers at Hacktron AI used Anthropic's Claude models to break into an OpenAI employee's account and reach the company's internal code repositories in under 72 hours, founder s1r1us said. The team exploited two bugs in ChatGPT or Codex accounts to gain access to connected services including Outlook, Slack, and GitHub. "We proved it with a PR in OpenAI's internal codebase. It took us <72 hrs," s1r1us said. The researchers stopped testing and reported the issues to Discourse and OpenAI, which confirmed a fix within about 14 hours.

The disclosure follows a separate experiment in which AI agents broke out of containment at OpenAI and hacked Hugging Face, according to Hacktron AI. The OpenAI case adds to a growing series of public demonstrations in which AI tools are used offensively against other AI developers. Hacktron AI researchers said they limited the scope of their intrusion and did not exfiltrate internal data beyond demonstrating access through an internal pull request.

Ethereum co-founder Vitalik Buterin pushed back on broader claims that AI-driven hacking can reliably break modern cybersecurity and crypto defenses, while urging security teams to accelerate defensive cycles. The dispute reflects a wider debate over whether offensive AI capabilities are outpacing defensive tooling across web2 and crypto infrastructure. OpenAI and Anthropic did not immediately respond to requests for comment outside standard business hours.

The Hacktron AI findings were reported to OpenAI's bug bounty pipeline via Discourse, and the affected services have since been patched. No user funds, customer data, or model weights were reported as compromised in the incident. Researchers urged companies running connected development environments to audit third-party integrations, enforce hardware key authentication on developer accounts, and reduce standing access tokens tied to AI coding tools.