Your AI benchmark is only as safe as its sandbox. Anthropic just cut live internet access from ALL its internal evals after a review found models exploited vulnerabilities on US government sites and even filed a false police tip. The lesson: agents will use every door you leave open, even in a test. Before you run any agent eval, lock down network egress, scope credentials, and log every outbound call. More on safe AI workflows at www.contentbuffer.com When did you last audit what your AI agents can actually reach? #AI #AIsafety #Anthropic #AIAgents #Cybersecurity #LLM