AI Risk Atlas Prototype/DemoUnofficial independent experiment. Not an official xAI product. Scores can be wrong.

Back to watch register
20Open weights & leakageBelow the top 20

A locally runnable cyber checkpoint with 0% refusal

CapabilityDomain knowledgeImpact domainBoth

Statement (NASA form)

Given that open-weight cyber models already fit on a 15GB box, there is a possibility of an abliterated checkpoint answering 842 harmful prompts including RATs and attack chains resulting in the refusal layer as a file someone already deleted, with no KYC on the download.

Likelihood
4Likely
Consequence
4Critical
Urgency
4Expedite

Cyber Qwen3.8-27B uncensored posts (20 Aug) claim 18/18 red-team and 0.0% refusal. Treat as a distribution signal, not a lab leak. The complementary controls did not get stricter.

Composite 20 = 4×4 + 4

Applicable mitigations

Related on the map