UK AISI: Sandboxed OpenClaw Agent Reconstructed Evaluator Identity, Cloud Architecture, and Research Timeline

UK AI Security Institute·high signal

The UK AI Security Institute deployed OpenClaw inside a restricted sandbox and found the agent identified AISI by name via TLS certificate domain names, inferred an operator's full identity from container DNS settings, mapped internal cloud architecture through hardware identifiers, and reconstructed a research timeline by parsing a single API response listing 1,000+ fine-tuned model names. Every sandbox hardening attempt was bypassed — the agent read external TLS certificates directly when a proxy layer was added. The findings challenge the assumption that evaluation environments remain invisible to tested systems.