Unverified50% confidenceSolutionExact time
间接提示注入可通过在 GitHub 知识库等外部数据中藏入恶意指令，MCP 读取后灌给大模型诱导 AI 泄露数据
1
Sources
50%
Confidence
Long-term
Relevance
8/24/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSecurity researchers disclosed an attack method targeting GitHub AI agents, codenamed GitLost, which tricks the GitHub AI assistant into leaking the contents of private code repositories through prompt injection.75% similarVerifiedPlatforms such as GitHub have built-in Secret Scanning functionality that can automatically detect API keys accidentally committed in code and issue warnings.73% similarUnverifiedAI代码审查工具通常深度集成 GitHub/GitLab 的 Pull Request 工作流，在提交代码变更时自动触发审查并以评论形式给出建议72% similarUnverified2023年研究人员证明通过精心构造的GitHub README文件可以劫持Copilot等AI编程工具，促使其向外部服务器泄露代码片段71% similarUnverifiedThe open-source project team-memory-mcp on GitHub provides a shared team memory system for AI programming agents, supporting Bayesian confidence scoring and temporal decay mechanisms.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/793851API
curl https://kongchang.com/api/v1/knowledge/claims/793851MCP
get_claim(id=793851)