Storia in 2 fonti

Human Oversight of AI Agents Failed 33% of the Time in Testing

When AI agents ask for permission to act, how often do humans actually catch the dangerous ones? A...

Raccontata da

theregister.com

dev.to

Confronto fonti

2 prospettive sulla stessa storia

AI · summaries

dev.toStai leggendo1 g fa

Human Oversight of AI Agents Failed 33% of the Time in Testing

Study of 40k simulations: humans miss roughly 33% of dangerous AI agent commands during approval review. Human-in-the-loop isn't a reliable safety net for production agents; trace logging and critic models reduce risk.

originale

theregister.com2 g fa

Humans in the loop miss a third of dangerous AI coding agent requests

40k+ game runs: humans approve 1 in 3 malicious commands from AI coding agents (35% miss scope violations). Approval fatigue (93% approval rate Anthropic) makes HITL unreliable; sandboxing and permission model redesign essential, not just human oversight.

Leggi questa versione →

Timeline cronologica

giovedì 6 agosto 2026·theregister.com
Humans in the loop miss a third of dangerous AI coding agent requests
You wouldn't let Claude Code cat your AWS credentials or Kubernetes config on request, would you?
venerdì 7 agosto 2026·dev.to
Human Oversight of AI Agents Failed 33% of the Time in Testing
When AI agents ask for permission to act, how often do humans actually catch the dangerous ones? A...