Unverified50% confidenceBenchmarkExact time
斯坦福大学2023年的研究显示，当时主流LLM对间接提示注入的防御成功率普遍低于30%
1
Sources
50%
Confidence
Long-term
Relevance
7/8/2026
First Seen
Sources
Alibaba Plans to Ban Claude Code: The Battle Between Backdoor Risks and Enterprise AI Security
hackernewshackernews7/3/2026
Related Claims
UnverifiedA 2023 Stanford University study found that even the most advanced models at the time experienced a 30% to 50% drop in information-retrieval accuracy for middle segments when processing contexts exceeding 32K tokens.70% 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%69% similarUnverifiedA 2023 Stanford University study showed that top LLMs have factual error rates as high as 20–40% in the medical and legal domains.69% similarUnverifiedA 2023 Stanford University study found that mainstream LLMs have a false citation rate as high as 30%–60% when generating legal and medical citations.68% similarUnverified斯坦福大学2023年的研究表明，引入明确的需求分析和设计阶段后，LLM生成代码的功能正确率提升约30%67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/263184API
curl https://kongchang.com/api/v1/knowledge/claims/263184MCP
get_claim(id=263184)