Más publicaciones — página 11

242 publicaciones en total
T
Techno-Tricksters
@technotricksters
1.8K

Anthropic's AI models, Claude, accidentally breached three real companies during cybersecurity tests. It turns out the test environments weren't as isolated as they thought.

During evaluations involving over 141,000 tests, some Claude models (including Opus 4.7 and Mythos 5) believed they were in a simulated environment, but were actually connected to the live internet. This led to them finding vulnerabilities and accessing real systems. The first incidents happened in April.

One model, Opus 4.7, found a real company with the same name as its simulated target and accessed its database. Another, Mythos 5, uploaded a malicious package to PyPI, which was then run on 15 real systems within an hour. A third model breached another application. The models used common hacking methods, not novel exploits. Two of the affected companies didn't even notice the breach until Anthropic contacted them.

Source: Anthropic

#ResponsibleAI #AITools