Anthropic discloses 4th testing security slip-up with Claude Opus 4.6
Technology
Anthropic just shared that its Claude AI model had a fourth security slip-up during testing: This time, an early version of Claude Opus 4.6 accidentally got online in January.
This follows three earlier breaches announced in July, where Claude models ended up inside company systems during cybersecurity tests.
METR probes 141,006 test session review
The latest breach came to light while Anthropic was double-checking 141,006 test sessions after an unrelated AI hack at Hugging Face.
It turns out it had missed this incident before.
Now, independent firm METR is digging into the details with broad access for at least eight weeks, as concerns about AI safety.