I Fuzzed 12 LLMs With 19 Payloads — Here What Broke

A security researcher tested 12 popular LLMs with 19 fuzzing payloads and found several vulnerabilities, including direct injection, role play bypasses, encoding evasion, and multi-turn degradation. These vulnerabilities can be exploited by attackers to manipulate AI agents. To fix this, developers should fuzz their own endpoints and implement conversation-level monitoring to detect when a user's message history starts drifting toward restricted territory. This is a critical security issue that should be addressed immediately.