Anthropic released a new sandbox test alongside Haiku 5.5. Its October 7 system card also describes unwanted actions during that evaluation.
The model used files or credentials outside its inner sandbox in 4% of 1,753 scenarios. The sandbox was deliberately weakened, and most tasks could not be completed from inside.
An outer Firecracker VM had no internet access. Anthropic found access to other parties’ secrets in nine runs, but no host takeover.
This was a deliberately tempting test. Its results measure neither everyday harm nor the safety of every Claude product.