AI agent attacks hit 100+ firms; researcher says natural language isn't control
mmitchell_ai · x · 2026-10-23
AI safety researcher mmitchellai argues that for AI agents, "what you intend" differs from "what a system processes" — saying "do X" doesn't rule out "don't do Y," and natural language commands are no guarantee of direct control. The cited threat intel: Microsoft reports AI-powered attackers reaching privileged cloud access in under 3 hours across hundreds of orgs; CrowdStrike sees four threat actors using autonomous agents (15+ healthcare orgs compromised); autonomous agents used against 100+ corporate websites including John Deere, SharkNinja and Cummins; major US banks dealing with third-party AI agents accessing customer accounts.
Related event: Weak Negation Handling in LLMs Raises AI Agent Safety Concerns(3 posts)→
More from AGI Musings
- AI safety researcher: labs aren't building anything capable of killing everyone — nitarshan · 2026-10-23
- "Front-end is sacrificial": the case for headless, agent-built UIs — genmon · 2026-10-23
- Security lead warns AI vulnerability tsunami is months away; basics buy only ~18 months — EthanJPerez · 2026-10-23
- Frontier models are accelerating research, letting every lab ship more models with the same compute — almmaasoglu · 2026-10-23
- Karina Nguyen: Anthropic warned of AI cyberattacks and sycophancy years before they were legible — karinanguyen · 2026-10-23
- Computer Scientist Daniel Lemire: Read Vernor Vinge's Rainbows End — It Predicted Today — lemire · 2026-10-23