Cyber toolkits for LLMs

Researchers from Anthropic and Carnegie Mellon University demonstrate that LLMs paired with a translator-style cyber toolkit named Incalmo can autonomously plan and execute complex, multistage attacks in high-fidelity simulated networks (25–50 hosts), fully compromising 5 of 10 test environments and partially compromising four more; results underscore how such tooling could lower barriers to complex cyber attacks while also offering potential to automate defensive testing. The experiments used known vulnerabilities in simulated environments without active defenses, so the work highlights capability risk and research implications rather than reporting active real-world exploitation.