[KongchangAI]
Unverified50% confidenceOpinionExact time

Hard-to-detect safety risks such as emergent misalignment need to be exposed in advance through attack-defense contests at the data level, rather than passively responded to after a model is deployed.

1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen

Sources

Related Entities

Related Claims

Cite This Claim

Stable URI
https://kongchang.com/claim/1133580
API
curl https://kongchang.com/api/v1/knowledge/claims/1133580
MCP
get_claim(id=1133580)