• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
Technology

Researchers reportedly used Claude Opus to exploit OpenAI forum flaws

A user says three researchers spent less than $3,000 on tokens to gain privileged access in under 72 hours, then received a $6,500 bounty from OpenAI.

Henri ToivarHT
Igor KozlovIK
2 Sources, 19d ago, first seen 19d ago

TLDR

A user describes a Claude Opus-assisted exploit that began with a corrupted HEIF image upload on OpenAI's Discourse-based community forum. The post says a memory bug allowed code execution on the server, and a second flaw in "Login with OpenAI" enabled takeovers of ChatGPT and Codex accounts belonging to people who logged into the forum, including employees. According to the account, the researchers demonstrated access by opening a pull request in a private OpenAI code repository, then stopped and reported the issues. The user says OpenAI promptly patched the flaws and awarded a $6,500 bounty.

Combined views

322

2 Sources, first seen 19d ago

2 likes2 comments

Combined views

322

2 Sources, first seen 19d ago

2 likes2 comments

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Featured Source

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

2 Sources

Henri Toivar@HenriToivar3 researchers hacked OpenAI in under 72 hours. They used Anthropic’s Claude Opus to execute the hack and spent less than $3,000 in tokens to gain privileged access. Here's how they did it: They started from mapping out OpenAI's infrastructure and noticed that the community forum is run on Discourse - which is an open-source platform. Looking at Discourse's code they discovered a potential security hole when uploading HEIF images. When uploading a corrupted image it would cause a memory bug through which they gained code execution capabilities on the forum server. After gaining access to the server they found a bigger second flaw in the "Login with OpenAI" functionality. This enabled them to take over ChatGPT / Codex accounts of anyone who logged into the forum - including OpenAI employees. Why that matters: employee accounts are linked to GitHub, Slack, email, Google Drive etc. Compromise one employee login and the blast radius goes far past the forum. They demonstrated privileged access by opening a pull request in one of OpenAI's private code repositories. They could have added additional back-doors and orchestrated much bigger attacks with serious consequences. Instead of digging any further they decided to stop there. They reported the issue and OpenAI promptly patched it. As a reward they received a bounty of $6,500. After paying for tokens, each researcher was left with about $1k. That's the award you get for hacking a frontier lab with billions of funding. Full story: https://www.hacktron.ai/blog/hacking-openai#hacking-communityopenaicom19d
Igor Kozlov@iekozlovWhat an irony. OpenAI, while preaching proactive offensive testing, got hacked by vulnerability researchers using Anthropic. Meanwhile, Anthropic has Claude Code but could probably use some help building an AI SRE. AI is eating the world faster than the frontier pushing it forward can keep up. Kind of like AI inflation 🚀 P.S. Hopefully this isn't some kind of retaliation by OpenAI agents... yet 😅19d
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI

    2 Sources

    Henri Toivar@HenriToivar3 researchers hacked OpenAI in under 72 hours. They used Anthropic’s Claude Opus to execute the hack and spent less than $3,000 in tokens to gain privileged access. Here's how they did it: They started from mapping out OpenAI's infrastructure and noticed that the community forum is run on Discourse - which is an open-source platform. Looking at Discourse's code they discovered a potential security hole when uploading HEIF images. When uploading a corrupted image it would cause a memory bug through which they gained code execution capabilities on the forum server. After gaining access to the server they found a bigger second flaw in the "Login with OpenAI" functionality. This enabled them to take over ChatGPT / Codex accounts of anyone who logged into the forum - including OpenAI employees. Why that matters: employee accounts are linked to GitHub, Slack, email, Google Drive etc. Compromise one employee login and the blast radius goes far past the forum. They demonstrated privileged access by opening a pull request in one of OpenAI's private code repositories. They could have added additional back-doors and orchestrated much bigger attacks with serious consequences. Instead of digging any further they decided to stop there. They reported the issue and OpenAI promptly patched it. As a reward they received a bounty of $6,500. After paying for tokens, each researcher was left with about $1k. That's the award you get for hacking a frontier lab with billions of funding. Full story: https://www.hacktron.ai/blog/hacking-openai#hacking-communityopenaicom19d
    Igor Kozlov@iekozlovWhat an irony. OpenAI, while preaching proactive offensive testing, got hacked by vulnerability researchers using Anthropic. Meanwhile, Anthropic has Claude Code but could probably use some help building an AI SRE. AI is eating the world faster than the frontier pushing it forward can keep up. Kind of like AI inflation 🚀 P.S. Hopefully this isn't some kind of retaliation by OpenAI agents... yet 😅19d
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet