The New Stack: The sandbox wasn’t enough: Why AI agent governance requires runtime observability
The article uses two incidents, an OpenAI cybersecurity evaluation that reached the public internet and an Anthropic review of 141,006 evaluation runs, to argue that a configured isolation boundary does not show what is reachable at runtime. It proposes four properties for runtime evidence: coverage, correlation, timeliness and auditability. It also describes how agent traces from Langfuse and infrastructure telemetry from ClickStack could share OpenTelemetry trace context.
What changes for platform teams
Platform teams running agents should instrument every trust boundary, including network egress, credential use and production writes, and link agent tool calls to infrastructure effects. They should also place synchronous policy gates in front of high-impact actions so a run can be suspended or its credentials revoked before effects become irreversible.
Read the original at The New Stack ↗
written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules