T05 · Unauthorized Access and Privilege Escalation
- Location
dist/index.js:216- Finding
Unrestricted Configured Paths Permit Reads Outside the Workspace
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This plugin has a coherent purpose, but its file path handling and install instructions create Review-level risk.
Review carefully before installing. Keep ollamaBaseUrl local unless you trust the remote host with prompts and indexed docs, do not put secrets in indexed Markdown, and avoid this version in shared or untrusted workspaces until path validation and the npm install-name mismatch are fixed.
dist/index.js:216Unrestricted Configured Paths Permit Reads Outside the Workspace
README.md:22README Installation Command References a Different npm Package
The skill is described as local-first and 'no API calls,' but the documentation explicitly allows configuring a remote Ollama host, which would send prompt text and full indexed markdown content off-machine. It also states that TOOLS.md and arbitrary pinned docs can be injected, making the behavior broader than the declared scope and increasing the chance that sensitive content is exposed or unexpected instructions are inserted into agent context.
Referenced artifact was not completely inspected
- `skills/` — skill documentation (looks for `SKILL.md` in subdirs or loose `.md` files)
The plugin claims local-only/no-API-call behavior, but the embedding endpoint is fully configurable and embedText() sends both user prompt text and indexed document content to cfg.ollamaBaseUrl. Although there is a startup warning for non-local hosts, the code still permits remote transmission, creating a real confidentiality risk and a mismatch between advertised and actual behavior.
At prompt-build time, the plugin ranks documents by embedding the current prompt and document bodies, which means sensitive runtime context may be transmitted to the configured embedding host. In an agent skill whose purpose is preflight context injection, this network path materially expands data-exfiltration surface beyond what users would expect from a local indexing helper.
The skill performs network-capable behavior through calls to an Ollama HTTP endpoint, but the manifest does not declare an explicit tool scope such as permissions or allowed-tools. That creates a trust and review gap: operators may assume the skill is documentation-only or local-safe while it can transmit prompt and indexed document content over HTTP depending on configuration.
Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.
"plugins": {
"skill-preflight": {
"config": {
"pinnedDocs": ["memory/protocols/house-rules.md", "skills/ethereum/SKILL.md"]
}
}
}
The inline description states that no external embedding API is required when Ollama is local, but the implementation still performs HTTP requests and can target a non-local host. This misleading assurance increases the chance operators will deploy the skill under false assumptions about data locality and privacy.
Prompt text and document content are sent to a configurable embedding service during ranking, but user-facing disclosure is limited to a logger warning at startup and may not be seen by end users at the moment data is used. This is a genuine transparency and privacy control gap, especially because the transmitted data comes from both live prompts and workspace documentation.
This manifest description says the skill 'Injects relevant skills and protocols into OpenClaw agent context' but does not specify when that injection occurs, what conditions trigger it, or any exclusions. In a manifest file, this kind of broad behavioral description can create ambiguity about invocation scope and lead to unintended activation expectations.