Anthropic's Claude helped researchers reach OpenAI staff accounts in three days.
Three security researchers at Hacktron used Anthropic's Claude models to develop exploit code against OpenAI and reached several OpenAI employee ChatGPT accounts, The Register reported.
The researchers wrote that they went from the first finding to demonstrating they could reach an internal OpenAI repository "by opening a harmless pull request" in under 72 hours, and then "stopped any further testing". No internal code or sensitive data was taken.
OpenAI patched the flaw in about 14 hours and paid a $6,500 bounty. The company said: "Testing against the Discourse-hosted community.openai.com was explicitly excluded from our bug bounty program. The award recognizes the OpenAI-side finding, not the actions against Discourse."
Anthropic did not respond to The Register's requests for comment.
The models that helped write the exploit are generally available. OpenAI's own statement puts part of the work on a Discourse-hosted community site that its bounty programme did not cover.
If a researcher found the equivalent in your estate this morning, how many hours until it reached your security team, and how many until it was closed?
Sources
Our file on Anthropic
- 21 Sept
Anthropic let Accenture safety-test its AI from inside. Both expect to invest $1 billion.
- 20 Sept
OpenAI's agent platform was installed at 75 firms. 69% made it their main one.
- 20 Sept
Anthropic's AI led none of its own research in February. Now it leads 26%.
- 19 Sept
Anthropic ordered $44.6 billion of computing from a company earning $140 million.
- 17 Sept
Anthropic prepared one unreleased feature linking its AI to personal bank accounts.
For your next steering committee
- 1Ask when systems of this kind in your landscape were last patched, and who confirms it.
- 2Ask which finance and ERP accounts can be reached from outside, and how each sign-in is protected.
- 3Ask how long a restore took the last time it was tested.
Issue 01 of the ERP LEADERS brief puts a timeline, the three-number review and six questions on one page. Read issue 01
Welcome back. · Issue 01