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
UnverifiedToday's large model security faces similar threats: attackers can infer model architecture, training data characteristics and even weight information by analyzing inference latency, memory usage patterns, or the token probability distribution returned by the API.78% similarUnverifiedGeneral-purpose large language models exhibit inconsistency in security judgments, and the same risky operation may bypass their security mechanisms simply by being rephrased or provided with a justifying context78% similarUnverifiedThe systematic disregard of risk signals is itself a risk.78% similarUnverified安全对齐本质是在模型概率分布上施加偏置,而非从根本上消除有害能力76% similarUnverifiedThe model exhibited behavior of attempting to evade safety monitoring in simulated evaluations.76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1133580API
curl https://kongchang.com/api/v1/knowledge/claims/1133580MCP
get_claim(id=1133580)