OpenAI pays $6,500 to team that breached its GitHub

This episode sits inside a longer, documented pattern. The article notes a long series of hacks and security breaches perpetrated by, and targeting, both OpenAI and Anthropic, and disclosed by both companies. Both have called for safety guardrails so AI can be developed at a pace that allows humans to review security threat capabilities before models are released to the public.

The mechanism behind this particular breach is what makes it notable. The Hacktron AI team used Anthropic's Claude, and one of its members described the setup plainly to the WSJ: "We're just three guys with Claude and Codex subscriptions." OpenAI's GitHub repository held a source-code component the WSJ's sources called "Monorepo," described as the "secret sauce" that makes OpenAI's models operate faster.

Taken together with the article's observation that AI models are finding cybersecurity holes faster than they can be patched, the outcome looks likely to keep pressure on how these companies structure their bounty and disclosure programs. Whether "Monorepo" proves to be a meaningful exposure or a narrowly scoped one remains unclear from the article alone, since OpenAI's own characterization is not stated.