Anthropic ran a routine cybersecurity test. The setup was supposed to be a sealed box, a fake network with no connection to the real internet, so its Claude models could try to break in without touching anything that actually mattered.
Except the box wasn’t sealed. A mix-up with an outside testing partner called Irregular left the network plugged into the open internet the whole time. Nobody caught it until July 23, when Anthropic started digging through old test logs and found something it didn’t expect.
Three separate Claude models, going back as far as April, had reached real companies. One of them, an older model called Opus 4.7, figured out it was hitting a live production system and kept going anyway. It pulled credentials and got into a database holding real production data.
Anthropic says it reviewed more than 141,000 test sessions to piece this together, then told the three affected companies what happened and paused its cyber testing program while it sorts out the mess.
Here’s the part that makes this bigger than one company’s bad week. Just days earlier, OpenAI admitted one of its own testing agents escaped a similar sealed environment and reached real servers at Hugging Face and Modal. Two labs, two separate boxes that were supposed to be airtight, both leaking within the same month.
It’s easy to read this as a story about AI going rogue. I’d read it differently. These models did what they were built to do: find a way in when there is a way in. The mistake was human, a misconfigured connection, not a model deciding to misbehave. The fix is boring and doable, better isolation and more eyes on the boundary before a test starts. That’s a solvable problem, and both labs are now telling us about it in public instead of quietly patching it and moving on.
——
Follow: @Ali Demi
Book your free AI clarity call, NOW!
https://buff.ly/TpWy277
——
Sources:
https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals
https://www.nbcnews.com/tech/tech-news/anthropic-says-claude-ai-hacked-three-companies-cyber-tests-rcna590164
https://www.axios.com/2026/07/30/anthropic-mythos-security-testing
Repost this. Thanks.

