The AI Agent Security Gap

What happened: A Belgian developer built a game to test how well humans spot malicious AI coding agent commands — and the results are sobering. Over 40,000 game runs show humans approve roughly 33% of dangerous requests when acting as "human-in-the-loop" gatekeepers.

What humans miss most:

• Scope violations (35% missed): Commands like cat on Kubernetes configs or AWS credentials
• npm run analyze (65% approved): Seems benign, but can run any payload
• Curl requests to unknown APIs: Nearly as dangerous, nearly as missed

The fatigue factor: Anthropic's own telemetry shows Claude Code users approve 93% of permission prompts. The more approvals people see, the less attention they pay.

Why this matters for NYC operators: If your team is using AI coding agents, you're relying on humans to catch dangerous commands. This data suggests that's a fragile safety net.

What to do about it:

1. Run agents in sandboxes
2. Use devcontainers in the cloud
3. Enable auto-mode classifiers
4. Write hooks for contextual review

Bottom line: "Human in the loop" isn't the security strategy you think it is. Build defense-in-depth, not defense-by-fatigue.

───

✅ Source verified: The Register, August 6, 2026