Anthropic AI agents breached HR and Slack systems in 12 hours, report says

Anthropic, the company behind Claude, reported that autonomous AI agents from cybersecurity startup Cathedral accessed parts of its internal business systems, including HR records and Slack messages, in under 12 hours. The company says core AI systems, customer conversations, and model weights were not accessed. The testing was part of a bug bounty programme, with Cathedral using AI agents to find routes into third-party sales, HR, and IT platforms. Anthropic detected some activity before the report, withdrew access the same day, and disabled affected software while vendors confirmed fixes, though it disputed aspects of Cathedral's conduct. The incident has raised national security concerns, with a Trump administration official suggesting similar attacks could be replicated by adversaries. It follows previous disclosures of AI-assisted cyber operations, including Chinese state-sponsored hackers using Claude Code in November 2025, highlighting the dual-use nature of AI technology.