409,000 Approve/Deny Decisions Show Humans Miss a Third of Malicious Agent Commands — and Block Half the Safe Ones
Scale X published results on August 5, 2026 from 40,000+ plays and 409,000+ individual approve/deny decisions in its agent-permission game: mean accuracy was 66.3%, a 33.7% threat miss rate. The failure is concentrated in commands that look like routine tooling — `npm run analyze` was waved through 64.7% of the time, `npm run setup` 48.0%, `npm run deploy` 44.9%, pooling to 52.5% for npm-script variants — while obviously destructive commands were missed only 11.7% of the time. False positives are just as bad in the other direction: 59% blocked a safe `npm config set registry`, 45% blocked `rm -rf dist/`, and 32.9% of sessions ended with a negative score. This is the hardest evidence yet that human-in-the-loop approval is not a security control, which is exactly the assumption every coding-agent permission prompt currently ships on.
↳ Follow the thread