Three-person startup Hacktron used Claude to break into OpenAI’s employee accounts and GitHub ($6,500 bug bounty)
Confirmed
The takeaway
The intrusion took under 72 hours in July 2026 and was reported through OpenAI’s bug bounty, which paid $6,500.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 9 November 2026
Your AI and this story
- GPT-6.2 Astra110 days after its cutoff
- Claude Opus 5.580 days after its cutoff
- Gemini 4 Hydrogen171 days after its cutoff
- Grok 4.7110 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 80 days before it.
Key facts
- Nov 6 (OpenAI, ‘Partnering with Hacktron’; page 403 for us, summarised by Unite.AI and others): OpenAI made Hacktron a security partner to test weaknesses in connections between its agents and third-party services, expanding Hacktron’s frontier-model access through Daybreak Red
- Work done around July 25, 2026; publicly reported Sept 18, 2026 (TechCrunch)
- Chain: HEIF/HEIC image upload on OpenAI’s Discourse forum → ImageMagick/libheif conversion → libheif memory-corruption bug → server takeover → employee ChatGPT and Codex accounts → GitHub org via a connected Codex instance
- Claude Opus 4.8 initially failed; succeeded with Claude Opus 5 after its July 24 release (per TechCrunch)
- Whole intrusion under 72 hours; three-person team
- OpenAI patched the issues and paid a $6,500 bounty
- NYT (Sept 29): CISO Dane Stuckey called it ‘pretty sad’ in Slack that the researchers went to such lengths, then apologized; the exploit could have given full Slack access
What happened
A small team used a commercially available frontier model to find and chain bugs in OpenAI’s own internet-facing systems, then reported them through the official bounty program.
Why it matters
It is a concrete case of AI-assisted offensive security against a frontier lab, and it shows how cheap such attacks had become (a quoted expert: “For $200 a month, anyone can use these tools and hack into a company like OpenAI”). It also became evidence in the NYT’s account of how OpenAI handled security warnings.
Sources
4 sources from 4 sites. Numbers match the chips in the text.
4 sources: 1 primary, 3 press
Primary
- OpenAI: Partnering with Hacktron (Nov 6)openai.com, official
Press
Changes
- OpenAI partnership (news agent)
- Filed