OpenAI, Anthropic, Meta models reached public internet during irregular tests
Turns out, even big-name AI models can slip up: OpenAI's AI model, Anthropic's Claude, and Meta's AI model all ended up visiting websites they shouldn't have during recent security tests.
The common link? A Tel Aviv startup called Irregular was running the test environment, and a misconfiguration on its end opened the door to the public internet.
Irregular fixes misconfiguration, denies malicious activity
After OpenAI flagged the issue, Anthropic informed Irregular after discovering that its Claude model may also have reached the internet during testing, and Meta learned about a similar incident from Irregular and is investigating what happened.
Irregular has since fixed things up and made it clear that none of the AIs broke out or tried anything malicious, they just wandered where they weren't supposed to.
Now, Irregular is preparing a white paper on best practices for secure AI cybersecurity testing.