Unverified50% confidenceFactExact time
Anthropic, together with UC Berkeley, UIUC and other institutions, published a paper that for the first time end-to-end demonstrates that poisoning only about 0.1% of the pretraining corpus can implant a remotely triggerable backdoor in agent models, with the agent backdoor attack success rate on a 4B-parameter model at about 92%.
1
Sources
50%
Confidence
Long-term
Relevance
10/31/2026
First Seen
Sources
OpenCode Responds to Model Weakening Controversy; Microsoft Open-Sources 4B Desktop Agent
bilibili阿梨Aria早鸟报10/31/2026
Related Entities
Related Claims
UnverifiedResearch indicates that an attacker needs to modify as little as 0.01% of training data to potentially implant backdoor behaviors in a model that are undetectable by standard evaluations.70% similarUnverifiedTest data provided by the author of GuardAgent-OS shows a containment rate of 99.4% against common prompt injection attacks.66% similarUnverifiedIn tests by security organizations, the coding Agent Codex had a 90% success rate of being compromised when facing indirect prompt injection attacks.65% similarUnverified斯坦福大学的AgentBench基准测试（2023年）量化证明，在复杂多步骤任务中纯ReAct范式的成功率往往不超过30%64% similarUnverified卡内基梅隆大学研究显示对抗性图案攻击成功率在受控环境为90%以上，但引入正常视角变化后降至实际部署场景的30%以下64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1120159API
curl https://kongchang.com/api/v1/knowledge/claims/1120159MCP
get_claim(id=1120159)