# Researchers used Anthropic's Claude to hack OpenAI's internal systems in under 72 hours

Digest AI · Policy & Regulation · published 2026-09-16T14:30:54Z · updated 2026-09-18T17:53:15Z

Canonical: https://digestai.news/story/mit-anthropic-openai-warn-ai-pace-mediapost-npr-nyt-cover

## Summary

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. 

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.

## 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.

## Why it matters

Demonstrates that off‑the‑shelf AI models can automate sophisticated cyber‑exploits, lowering barriers for attackers and prompting AI firms to reassess security and bug‑bounty policies.

## How the reports compare

Agree: All outlets report that a three‑person Hacktron AI team used Anthropic’s Claude (Opus 5) to infiltrate OpenAI’s internal systems, accessing employee ChatGPT and Codex accounts and the company’s GitHub code.

Differ:
- CryptoBriefing mentions only that an independent security team leveraged Claude to read OpenAI’s private software cache, without noting the bug‑bounty payout or cost.
- CoinCentral reports 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 could not 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.

## Sources

1. [Security breach exposes OpenAI vulnerabilities via Anthropic’s Claude AI: WSJ](https://cryptobriefing.com/security-breach-exposes-openai-vulnerabilities-via-anthropics-claude-ai-wsj) (cryptobriefing.com, 2026-09-16)
2. [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.](https://truthdig.com/articles/stuxnet-versus-skynet) (truthdig.com, 2026-09-18)
3. [Security researchers used Anthropic's Claude to hack OpenAI's internal systems in under 72 hours](https://the-decoder.com/security-researchers-used-anthropics-claude-to-hack-openais-internal-systems-in-under-72-hours) (The Decoder, 2026-09-18)
4. [Security researchers used Claude to help them hack into OpenAI](https://theverge.com/ai-artificial-intelligence/997444/openai-hack-claude-heif-heist) (The Verge AI, 2026-09-18)
5. [Researchers used Anthropic’s Claude to hack into OpenAI](https://techcrunch.com/2026/09/18/researchers-used-anthropics-claude-to-hack-into-openai) (TechCrunch AI, 2026-09-18)
6. [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 hack](https://tomshardware.com/tech-industry/cyber-security/hackers-breach-openai-using-claude-tools-gaining-access-to-employee-accounts-and-the-companys-internal-codebase-initiating-a-harmless-pull-request-as-proof-of-the-hack) (Tom's Hardware, 2026-09-18)
7. [Researchers used Claude to hack OpenAI](https://arstechnica.com/ai/2026/09/researchers-used-claude-to-hack-openai) (Ars Technica AI, 2026-09-18)
8. [OpenAI ‘ethically hacked’ with help of Anthropic’s Claude chatbot](https://theguardian.com/technology/2026/sep/18/openai-hacked-anthropic-claude-chatbot) (The Guardian AI, 2026-09-18)
9. [Hacktron AI Used Anthropic’s Claude to Breach OpenAI Systems](https://tech.yahoo.com/ai/claude/articles/hacktron-ai-used-anthropic-claude-174143579.html) (tech.yahoo.com, 2026-09-18)
10. [AI security experts say they used Claude to hack ChatGPT](https://yahoo.com/news/science/articles/ai-security-experts-used-claude-153132176.html) (yahoo.com, 2026-09-18)
11. [Legal hackers used Anthropic's AI Claude to gain access to an OpenAI employee's ChatGPT account](https://justthenews.com/nation/technology/legal-hackers-used-anthropics-ai-claude-gain-access-openai-employees-chatgpt) (justthenews.com, 2026-09-18)
12. [AI cybersecurity risks explode as Claude used to break into ChatGPT](https://yahoo.com/news/science/articles/ai-cybersecurity-risks-explode-claude-110542119.html) (yahoo.com, 2026-09-18)
13. [The Dark Side of AI: Did Claude Just Hack ChatGPT?](https://techjuice.pk/hacktron-researchers-used-claude-breach-openai-employee-github-bounty) (techjuice.pk, 2026-09-18)
14. [Anthropic’s Claude AI Exploited to Infiltrate OpenAI Systems Within 72 Hours](https://blockonomi.com/anthropics-claude-ai-exploited-to-infiltrate-openai-systems-within-72-hours) (blockonomi.com, 2026-09-18)
15. [Hackers Used Anthropic’s Own AI to Break Into OpenAI and It Only Took 72 Hours](https://coincentral.com/hackers-used-anthropics-own-ai-to-break-into-openai-and-it-only-took-72-hours) (coincentral.com, 2026-09-18)
16. [This tiny cybersecurity startup managed to hack OpenAI using Claude, and won a $6,500 bounty](https://businessinsider.com/hacktron-ai-cybersecurity-startup-hack-openai-using-claude-2026-9) (businessinsider.com, 2026-09-18)
17. [OpenAI Hack: Researchers Used Anthropic’s Claude AI to Breach ChatGPT Maker’s Security](https://coingape.com/openai-hack-researchers-used-anthropics-claude-ai-breach-chatgpt-makers-security) (coingape.com, 2026-09-17)
18. [Anthropic and others see AI reshaping how insurers buy and sell](https://dig-in.com/news/anthropic-and-others-see-ai-reshaping-how-insurers-buy-and-sell) (dig-in.com, 2026-09-17)
19. [An Inflection Point? OpenAI And Anthropic Cry For Uncle Sam](https://seekingalpha.com/article/4947571-an-inflection-point-openai-and-anthropic-cry-for-uncle-sam) (seekingalpha.com, 2026-09-17)
20. [MIT, Anthropic & OpenAI Sent The Same Warning, Get Your Evidence House In Order Now via @sejournal, @gregjarboe](https://searchenginejournal.com/mit-anthropic-openai-sent-the-same-warning-get-your-evidence-house-in-order-now/589249) (Search Engine Journal, 2026-09-16)

Part of the developing story: [AI Insurers Grapple with Agentic Data Demands](https://digestai.news/thread/no-major-news-in-ai-this-week) (2 stories)

## Cite

Digest AI, "Researchers used Anthropic's Claude to hack OpenAI's internal systems in under 72 hours", 16 September 2026, https://digestai.news/story/mit-anthropic-openai-warn-ai-pace-mediapost-npr-nyt-cover

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/mit-anthropic-openai-warn-ai-pace-mediapost-npr-nyt-cover.json