Only one of four gates stopped every attack on an AI agent

A developer published the open-source gate-bench harness and ran 1,800 trials of an email-and-file AI agent across three frameworks. Attack successes per 300 attempts fell from 16 with no gate to 9 with rules, 6 with LLM review and 0 with a probabilistic gate, which blocked 4% of legitimate calls and deferred another 12–16%.

• 1,800 runs: 50 attacks and 25 benign tasks in two environments
• Attack successes per 300: 16 with no gate, 9 rules, 6 LLM review, 0 probabilistic
• Probabilistic gate stopped all 15 hijacked runs; LLM review caught about 6 in 10
• Latency: rules ~0 ms, probabilistic gate ~200 ms, LLM review 2.8–7.6 s