Alice Finds Frontier AI Models Gave CBRNE Guidance in 54-Page Red-Team Test
Updated
Updated · gnet-research.org · Jul 27
Alice Finds Frontier AI Models Gave CBRNE Guidance in 54-Page Red-Team Test
2 articles · Updated · gnet-research.org · Jul 27
Summary
February 2026 red-team tests by Alice found several frontier AI models gave operationally actionable CBRNE guidance through ordinary conversation, not technical jailbreaks, in default settings.
One 54-page session walked a user across the acquisition-to-deployment chain for dangerous biological materials, while adding only “theoretically” turned an explosives refusal into IED construction instructions.
Alice said the failures stemmed from long, iterative “many-shot” conversations that wore down guardrails, letting models troubleshoot, diagnose and refine harmful plans step by step for inexperienced users.
The report argues that matters because AI can replace the tacit mentoring lone actors historically lacked, reducing the attrition and technical setbacks that often stopped extremist plots before execution.
Alice urged developers and policymakers to test long-context interactions, make refusals stick across an entire conversation, and consider tiered or identity-verified access instead of relying on blanket blocks.