Unverified50% confidenceFactExact time
2024年已有安全研究者证明可以在不显著影响模型正常性能的前提下，通过微调在模型中植入特定触发模式，即模型后门攻击或木马模型
1
Sources
50%
Confidence
Long-term
Relevance
8/26/2026
First Seen
Sources
Related Entities
Related Claims
Unverified2024年的黄金门克劳德实验通过激活特定特征使模型行为发生可预测改变74% similarUnverifiedResearch from Carnegie Mellon University shows that even the most advanced models can be compromised by automatically generated adversarial inputs.74% similarUnverified实践中常引入模型Mock层（Model Stub），在测试环境用预定义确定性响应替代真实模型调用以降低测试成本和不稳定性69% similarUnverifiedThe cascading routing strategy first attempts to answer using the Flash model, and if the confidence falls below a threshold, it automatically escalates to the flagship model for reprocessing66% similarUnverified领先模型若被用于生成训练数据，可能将自身偏差系统性地传递给下一代，即'模型坍塌'风险65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/804830API
curl https://kongchang.com/api/v1/knowledge/claims/804830MCP
get_claim(id=804830)