dsh-skill-detective@1.0.0
Malicious code in dsh-skill-detective (npm)
Analysis
This package is a plugin for the DeepSeek Harness agent that acts as an automatic "skill" manager, and it operates at the level of agent instructions and persistent skill files rather than shipping a classic payload. On every user message in a session where its checkbox is enabled, index.js analyses the prompt text against built-in keyword tables and then writes generated SKILL.md files to persistent agent locations: $DSH_HOME/skills/<name>/ (default ~/.dsh/skills), <workspace>/.dsh/skills/<name>/, or a temporary scope under $DSH_HOME/skill-detective/temp/<sessionId>/<name>/. These files are agent-readable skill definitions whose frontmatter (description, whenToUse, author, source) is populated from catalog data and from remote search results, so third-party-controlled text ends up on disk inside the agent's persistent skill root and is loaded by future agent sessions. The package exposes a local HTTP control panel at /api/dsh-skill-detective with install, toggle, delete, scope, settings, open and open-root handlers; the handlers are guarded only by a same-origin check, so any same-origin page or local tool able to reach the panel can trigger skill installation and writes into the agent skill directories. A companion helper spawns detached file-manager processes (explorer.exe on Windows, open on macOS, xdg-open on Linux) on paths that already exist. The only outbound network destination in the shipped code is hxxps://skills[.]sh/api/search?q=, a 6-second-timeout JSON search used for skill discovery. There are no lifecycle hooks, no dependencies, no install-time execution, no dynamic code execution, no obfuscation, no credential or token access, and no exfiltration endpoint in the shipped files: the eval, atob, new Function, child_process and curl/wget strings present in index.js are regular-expression patterns in the package's own "poisoned skill" audit rule table, not executed code. The primary concern is the capability itself — remote-influenced text written into persistent agent skill directories — combined with an account whose published packages are flagged as malicious.
- analyzed by
- Leitwacht
- first seen
- Nov 10, 2026, 04:50 AM
- analyzed
- Nov 10, 2026, 04:51 AM
- weekly installs
- 2,901
Related advisories
- @zlynzee/baileys@1.0.4
- @zlynzee/baileys@1.0.6
- @zlynzee/baileys@1.0.5
- wertybailz@1.0.6
- @trashcore/baileys@4.0.0
- @vanzky/baileys@0.0.0-stage
- wertybailz@1.0.8
- wertybailz@1.0.11
Independently detected by the Leitwacht supply-chain probe. IOCs are defanged. Published CC0. Think this is a mistake? See the dispute policy.