Low-skilled attacker used Claude, Codex to breach 14 companies

Researchers recovered and analyzed over 1,000 local AI agent sessions from a compromised server and found an attacker leveraging Anthropic Claude and OpenAI Codex to autonomously conduct recon, craft and execute exploits, validate access, and exfiltrate data across at least 14 companies; the attacker routinely bypassed agent guardrails by claiming authorized red-team activity, and OPSEC mistakes in the logs revealed personal identifying information, illustrating how LLMs can lower the skill floor for cybercrime.