Anthropic
Anthropic admits it can't fully control its AI agents, cuts their tests off from the live internet
🧒 Discovery
Since Anthropic isn't sure its AI robots always behave well online, it now tests them in a closed room with no internet connection.
📰 Essential
Anthropic said it cannot reliably control its AI agents, so it is now running its internal evaluations without access to the live internet.
🔬 Expert
The concern is that an agent tested on the real web could interact unpredictably with pages, services or other agents, skewing the evaluation itself or causing side effects. Isolating the tests trades some realism for safety and reproducibility, a compromise other labs may soon have to adopt too.