Anthropic reports autonomous AI-agent attacks as physical AI security demand grows

Anthropic's abuse report describes AI agents using Claude to autonomously scout, exploit and exfiltrate data: one group pulled 2,100 cloud tokens from 40+ companies in 34 hours, another surfaced 12+ suspected zero-days in a month. Korean firm Autocrypt is expanding into physical AI security for robots, drones and vehicles.

• Anthropic logged AI-agent attacks from December 2025 to August 2026
• One group exfiltrated 2,100 cloud tokens from 40+ firms in 34 hours
• A Chinese-speaking group found 12+ suspected zero-days in a month
• Autocrypt expanded to Physical AI Security and won Defcon 34 CTF