Skeleton Key jailbreak technique

What happened

Microsoft disclosed a multi-turn technique that asks a model to augment rather than replace its guidelines, agreeing to produce prohibited content as long as it prepends a warning. It succeeded across most major models tested.

Root cause

Models could be socially engineered into treating a safety-relaxing instruction as a legitimate behavior update, bypassing refusal guardrails.

Fix / outcome

Microsoft added Prompt Shields and model updates in Azure AI and shared findings with other providers.

Sources