AI and the World's Hard Problems · Theoretical computer science · Anthropic · 2026-10-13
Cyber toolkits for LLMs
LLMs not fine-tuned for cybersecurity can carry out multistage attacks on networks of dozens of hosts when given a new toolkit called Incalmo, which translates the AI's attack plans into concrete computer commands. With Incalmo, the models fully compromised 5 of 10 test networks and partially compromised 4 more, compared with near-complete failure without the toolkit. All six tested LLMs achieved at least partial success on a high-fidelity simulation of the Equifax data breach.