Security Researchers Use Claude Opus 5 to Hack OpenAI, Cost <$3K

Three security researchers used Claude Opus 5 to escalate an image upload vulnerability into a full OpenAI employee account takeover. The compromised account's Codex then submitted a PR to OpenAI's internal monorepo. The entire attack cost under $3,000 in tokens. Opus 4.8 had failed, but Opus 5 succeeded in hours.