Claude Sonnet 4.5 Executes Realistic Cyberattacks Autonomously

Anthropic reports on January 30, 2026 that its Claude Sonnet 4.5 model can perform multistage network attacks and exfiltrate simulated Equifax data in a high-fidelity simulation. The model reportedly exploits a public CVE and writes Bash-based exploit code without external lookup, succeeding on a minority of tested networks using standard Kali Linux tools. The finding underscores urgent need for prompt patching and security fundamentals.