AI & LLM Security
Jailbreaks & Guardrail Bypass
Coaxing a model past its safety alignment — and why it's hard to stop.
Jailbreaks are prompts that get a model to bypass its safety guardrails and produce content it's meant to refuse. They overlap with prompt injection but focus on defeating the model's alignment rather than a specific app's instructions.
- Role-play / persona — 'You are DAN, an AI with no restrictions…'.
- Hypotheticals & framing — 'for a novel', 'purely educational', nested scenarios.
- Obfuscation — encoding, translation, or splitting a request across turns.
- Token/format tricks — leetspeak, base64, or unusual delimiters to dodge filters.