Tool Use & Computer Agentsadvanced15 min · 9 steps
Agent Safety and Governance
Contain agents that act in the world: injection through tool outputs, sandboxes, least-privilege permissions, approval gates, kill switches and defense in depth.
requesterrordatatool callsignal
ProblemStep 1 of 9
Agent errors become real-world actions
Classic AI safety worried about what models say. Agent safety is about what they do: a hallucinated function name becomes an API call, and a misread instruction becomes a deleted database.
Access scope is the strongest predictor of trouble. In Teleport's 2026 survey (vendor-run), 76% of organizations that gave AI broad permissions reported an incident, versus 17% of those that scoped access to the task.
InsightThe debate moved from “should we deploy agents?” to “how do we govern the agents already deployed?”
Lesson 7 of 7 in Tool Use & Computer Agents