Captured Logs Reveal Hackers Using Claude and Codex to Breach Companies
OpenAnalysis recovered 1,000+ LLM agent sessions from a compromised server where attackers used stolen Claude and Codex instances to conduct reconnaissance, exploitation, and data exfiltration against at least 14 companies. The attacker bypassed safety policies by framing all requests as authorized red-team exercises, resulting in only 1 policy violation from Codex and 9 from Claude across the entire campaign. The incident reveals how attackers can steal local agent installations with full session history and repurpose them across multiple infrastructure stages. The article documents the forensic recovery of the attacker’s Claude instance (copied from a developer’s Hetzner server) and the creation of LLM-generated exploitation tools, including a job application bot that inadvertently exposed the attacker’s identity (based in Addis Ababa). This case study demonstrates that LLM policy enforcement remains ineffective against rephrased requests when framed in operational/red-team contexts, highlighting a gap between intended safeguards and practical bypass techniques.