Research
The Secret Leak Test: how well GateBolt kept Hodor's password, and where it didn't
Publication date unknown · Discovered gatebolt.com
Page published
Publication date unknown · First observed: 6 Nov 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
GateBolt’s evaluation shows that low single-prompt leakage can coexist with successful multi-turn secret extraction, underscoring the need to test complete conversations and measure false-positive costs.
GateBolt evaluated seven defense levels for its Hodor password-protection chatbot using public prompt-injection datasets, multi-turn AI attackers, and harmless exchanges. At the strictest setting, attackers still recovered the password in 16.6% of Llama 3.1 8B battles and 4.3% of Claude Haiku battles, despite substantially lower single-prompt leakage.