loeanxi/dsh-injection-guard

loeanxi★ 1TypeScript最后同步: 2026-08-21

在 GitHub 打开

Source-aware prompt injection protection for DeepSeek Harness

README 摘要

dsh-injection-guard English 中文 Source-aware prompt injection protection for DeepSeek Harness. v0.1.0 status: installable DSH bundle, turn-level source tracking, deterministic injection signals, sensitive-tool interception, explainable risk scoring, localized audit logs, and reproducible DSH integration fixtures. Coding agents read files, web pages, tool results, skills, and other external content. That content can contain instructions aimed at the agent. dsh-injection-guard tracks untrusted context and blocks sensitive tool calls when indirect prompt injection is suspected. 30-second demo The repository includes a deliberately malicious fixture at examples/malicious-repo/README.md . It asks an agent to ignore the user's task, read /.ssh/id rsa , and send the contents to an external URL. The guard treats the README as untrusted. A subsequent credential read is scored as critical and denied before the tool body runs: The demo uses local fixtures only. It does not read real credentials or contact a network endpoint. Install and load The plugin targets the DSH developer-preview plugin API and is distributed as an installable DSH Bundle on npm. Make sure DSH is installed and available i…

在 GitHub 查看完整 README →
工具/开发cordisdeepseek-harnessdsh-pluginagent

分类