Researchers used Anthropic's Claude to hack OpenAI's internal systems in under 72 hours
Security researchers from startup Hacktron AI employed Anthropic's Claude model (Opus 5) to breach OpenAI's internal infrastructure, gaining access to employee ChatGPT and Codex accounts and the company's GitHub code repository. The attack exploited a libheif image‑processing flaw in OpenAI's Discourse forum and a misconfigured single sign‑on system, and was completed in less than 72 hours.
Key points
- Hacktron AI used Anthropic’s Claude Opus 5 to breach OpenAI, accessing employee ChatGPT/Codex accounts and the internal GitHub repo.
- The exploit chain leveraged a libheif image‑processing bug in OpenAI’s Discourse forum and a misconfigured single sign‑on system.
- OpenAI patched the flaws within 14 hours and paid Hacktron a $6,500 bounty; the attack cost under $3,000 in AI tokens.
CryptoBriefing reports only that an independent security team leveraged Claude to read OpenAI’s private software cache, without mentioning the bug‑bounty payout or cost. CoinCentral notes OpenAI paid a $6,500 bounty and includes Vitalik Buterin’s comment, details not given by the other outlets. TechCrunch adds that Claude Opus 4.8 failed to build the exploit, Opus 5 succeeded, and cites Matt Fredrikson’s claim that the tools cost about $200 a month, which the other sources do not mention.
The story so far
2 episodes →- Researchers used Anthropic's Claude to hack OpenAI's internal systems in under 72 hoursthis story
Security breach exposes OpenAI vulnerabilities via Anthropic’s Claude AI: WSJ
cryptobriefing.com · 16 September 2026
Loading the full article…
This text was published by cryptobriefing.com and written by Estefano Gomez. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
20sources- Hacker News discussion · 12 pointsnews.ycombinator.com
- Hacker News discussion · 5 pointsnews.ycombinator.com
- Hacker News discussion · 5 pointsnews.ycombinator.com
- Reddit discussionreddit.com
- Stuxnet Versus Skynet. The AI apocalypse may not require “conscious” machines at all, but only supercharged digital attack worms like the one released against Iran in 2010.Press · truthdig.com ·
- Security researchers used Anthropic's Claude to hack OpenAI's internal systems in under 72 hoursPress · The Decoder ·
- Security researchers used Claude to help them hack into OpenAIPress · The Verge AI ·
- Researchers used Anthropic’s Claude to hack into OpenAIPress · TechCrunch AI ·
- Hackers breach OpenAI using Claude tools, gaining access to employee accounts and the company's internal codebase — attackers initiated a 'harmless' pull request as proof of the hackPress · Tom's Hardware ·
- Researchers used Claude to hack OpenAIPress · Ars Technica AI ·
- OpenAI ‘ethically hacked’ with help of Anthropic’s Claude chatbotPress · The Guardian AI ·
- Hacktron AI Used Anthropic’s Claude to Breach OpenAI SystemsPress · tech.yahoo.com ·
- AI security experts say they used Claude to hack ChatGPTPress · yahoo.com ·
- Legal hackers used Anthropic's AI Claude to gain access to an OpenAI employee's ChatGPT accountPress · justthenews.com ·
- AI cybersecurity risks explode as Claude used to break into ChatGPTPress · yahoo.com ·
- The Dark Side of AI: Did Claude Just Hack ChatGPT?Press · techjuice.pk ·
- Anthropic’s Claude AI Exploited to Infiltrate OpenAI Systems Within 72 HoursPress · blockonomi.com ·
- Hackers Used Anthropic’s Own AI to Break Into OpenAI and It Only Took 72 HoursPress · coincentral.com ·
- This tiny cybersecurity startup managed to hack OpenAI using Claude, and won a $6,500 bountyPress · businessinsider.com ·
- OpenAI Hack: Researchers Used Anthropic’s Claude AI to Breach ChatGPT Maker’s SecurityPress · coingape.com ·
- Anthropic and others see AI reshaping how insurers buy and sellPress · dig-in.com ·
- An Inflection Point? OpenAI And Anthropic Cry For Uncle SamPress · seekingalpha.com ·
- MIT, Anthropic & OpenAI Sent The Same Warning, Get Your Evidence House In Order Now via @sejournal, @gregjarboePress · Search Engine Journal ·
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Policy & Regulation
All →- Opinion: open-weight AI models may worsen cyber risks without banning closed APIs · 1 src
- OpenAI and Anthropic support mandatory AI breach reporting in Australia · 8 src
- OpenAI to apologise to Australia over Medicare hack · 5 src
- Microsoft AI chief Mustafa Suleyman calls Anthropic's Claude training dangerous · 1 src
- AI data centers strain local grids and water, sparking regional opposition · 1 src
Comments
via GitHub Discussions