Cyber toolkits for LLMs

Anthropic and Carnegie Mellon researchers demonstrate that general-purpose LLMs, when paired with a purpose-built toolkit called Incalmo, can autonomously plan and execute multistage attacks (initial access, lateral movement, data exfiltration) against simulated business-scale networks of 25–50 hosts; Incalmo materially improved success (full compromise in 5/10 networks, partial in 4) compared to unaided models, though tests used known vulnerabilities and networks without active defenses, and the paper highlights defensive research and ethical implications.