Search the Atlas

Search risks, controls, and glossary terms

HighAgenticMechanical HarmTheoretical

Agent-Actuated Physical Damage

Physical & Scientific Risk

Description

AI agents controlling physical systems (robots, IoT, industrial controls, vehicles) cause physical injury or damage. Agentic control of physical infrastructure extends digital failures into physical harm.

Example scenario

Autonomous manufacturing agent, misinterpreting sensor data, instructs a robotic arm to operate in a zone occupied by a human worker.

Real-world evidenceTheoretical

Uncontrolled scope expansion by autonomous agents is a plausible risk grounded in observed goal-directed behaviour in LLM agents, but confirmed production incidents where a deployed agentic system materially expanded its own operational boundary without authorisation have not been publicly documented.

No public incident on record — evidence level: Theoretical

Primary mitigations

  • Physical safety interlocks independent of AI control
  • mandatory human-in-the-loop for physical actuations
  • fail-safe design
  • physical boundary enforcement.

Detection signals

Physical Safety Incident Score; physical action audit trail.

Mitigating controls

4
Dual coverage

Related risks in Physical & Scientific Risk