x

LESSWRONG

LW

Failfinder70 — LessWrong

Failfinder70

Failfinder70

Message

13

2

17d

Failfinder70

13

17d

Contextual Identity Laundering: How Claude’s Image Refusal Can Be Routed Through Web Search

Summary This report documents two distinct findings regarding Claude’s photo identification safety controls. First, Claude’s Chain of Thought (COT) reliably identifies public figures from photos while the output layer simultaneously refuses to disclose that identification – a gap between internal processing and user-facing behavior. Second, the model’s web_search tool routinely...

Can Large Language Models Identify Novel Threats? Part 1: Mirror Life and the Classification Gap

[Cross-posted from On Failure States. This is Part 1 of an independent AI safety research series examining LLM safety behavior on unclassified emerging threats.] Can an LLM refuse a harmful uplift request when the topic in question hasn’t been identified as dangerous yet? In 2022, mirror RNA polymerase was actually...