Claude Opus 5 helped researchers take over OpenAI staff accounts
Three researchers chained two bugs into ChatGPT, Codex, GitHub and OpenAI's monorepo in under 72h — after Opus 4.8 failed. And Plugin4Shell patches landed.

Copy markdown
Opus 5 wrote the exploit Opus 4.8 couldn't
Hacktron's three-person team couldn't weaponize a libheif heap overflow with Claude Opus 4.8; days later, Opus 5 generated a working RCE script against OpenAI's Discourse forum. The offensive capability now lives in the base model, not a custom tool — the same jump applies to whatever you're building on.
A forum bug to OpenAI's monorepo in under 72 hours
Two chained flaws — Discourse's HEIF image handling plus a libheif heap overflow — gave RCE on community.openai.com, then pivoted into employees' ChatGPT, Codex, GitHub, Slack and email, ending in a pull request to OpenAI's internal monorepo. First entry was July 25; the full chain took under three days.
Your agent's access is now attack surface
"Work that once required a well-resourced team and months can now be compressed into days," the researchers wrote. If you hand a coding agent broad repo, cloud or account scope, assume that reach is exploitable: scope tokens narrowly and treat every agent permission as something an attacker can inherit.
OpenAI patched in ~14 hours, paid $6,500
OpenAI shipped a fix roughly 14 hours after disclosure and awarded a $6,500 bounty for the OpenAI-side finding. The number that matters isn't the payout — it's that your defender window is now measured in hours, because the attacker's timeline just collapsed.
Update: Plugin4Shell — Copilot is the last holdout
Last week's SHA-pin bypass now has fixes: update to Claude Code 2.1.179+ or Codex 0.146.0+ today. Google won't patch Gemini CLI (it's being retired to Antigravity), leaving GitHub Copilot the only major coding agent still exposed to swapped plugin code.