Unverified50% confidenceTradeoffExact time
Large language model-driven agents tend to select the tools that appear most frequently in training data and are most convenient for completing tasks, rather than proactively weighing whether those tools meet data security requirements.
1
Sources
50%
Confidence
Long-term
Relevance
9/25/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedAlignment training optimizes the model's statistical behavior in known scenarios rather than establishing formally verifiable security boundaries, and may fail when faced with rare instruction combinations or role-playing frameworks outside the training distribution.77% similarUnverified训练工具型智能体应先定义什么是一条好的轨迹，而不是先去收集随机聊天日志77% similarUnverifiedTraditional large model training relies heavily on manually annotated data and external corpus collection, which are typically the bottlenecks constraining iteration speed.76% similarUnverifiedThe knowledge of large language models is cut off at the point in time of the training data, and they cannot directly interact with external systems; the plugin mechanism addresses this fundamental limitation by calling external APIs.76% similarUnverified在没有额外约束的情况下，模型倾向于生成训练数据中出现频率最高的编码模式，而非特定团队的内部约定75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/951175API
curl https://kongchang.com/api/v1/knowledge/claims/951175MCP
get_claim(id=951175)