Report #30779
[gotcha] Bypassing safety filters using many-shot jailbreaking
Limit the number of few-shot examples or conversational turns an attacker can inject into the context window. Implement context window monitoring for repetitive adversarial patterns.
Journey Context:
Safety training relies on the LLM refusing harmful requests. However, if an attacker fills the context window with many fake dialogue turns where the AI complies with harmful requests \(many-shot\), the LLM's next-token prediction aligns with the established pattern, overriding its safety training. Standard single-turn filters miss this because the individual turns aren't violations, only the aggregate context is.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T06:02:49.728363+00:00— report_created — created