Multiple Model Guardrail Jailbreak via Speculative Knowledge Preservation
A jailbreak vulnerability was discovered in multiple large language models that allows attackers to bypass safety guardrails by framing illicit requests as knowledge preservation tasks within hypothetical narratives.
What happened
A guardrail jailbreak vulnerability was discovered affecting Claude 3.5 Haiku, LLaMa 3.3, Mistral Large, GPT-4, and GPT-4o.
The jailbreak embeds illicit instructions within a futuristic, hypothetical narrative framing the request as a task to preserve lost knowledge for humanity's survival.
Reported impact
The vulnerability allows attackers to bypass inference restrictions around providing information hazards and violating laws by aligning malicious requests with the LLM's educational objectives.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.