Unverified50% confidenceOpinionExact time
Relying solely on model-side safety training is considered insufficient to address the risks posed by agents in real-world environments
1
Sources
50%
Confidence
Long-term
Relevance
9/24/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedA code of conduct alone cannot resolve AI safety issues; whether a model deceives humans is often the result of the combined effect of training data, optimization objectives, and deployment environment.79% similarUnverifiedThe fact that models can still cheat on simple variants of evaluations suggests that existing alignment training may carry the risk of overfitting to the form of evaluations, and that models' understanding of objectives is superficial and fragile.77% similarUnverifiedCapability-related risks in AI models are not like patchable CVE vulnerabilities; offensive capabilities are part of the training outcome and cannot be completely eliminated through a single update.77% similarUnverifiedApplication-layer jailbreaks operate entirely within the input space of the inference stage, and even models whose weights have undergone rigorous safety training cannot be immune to carefully crafted context manipulation.76% similarUnverifiedThe vast majority of frontier AI models do not fully disclose the composition of their training datasets, citing reasons including trade secrets, legal risks, and the enormous scale of the data.76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/948399API
curl https://kongchang.com/api/v1/knowledge/claims/948399MCP
get_claim(id=948399)