delegate-explore
Pass
Audited by Gen Agent Trust Hub on Oct 24, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from untrusted external sources including the public web (WebSearch/WebFetch) and third-party enterprise platforms (Notion, Confluence, Jira) via MCP tools. This data is processed by subagents, creating a surface for indirect prompt injection where malicious instructions embedded in external content could influence the subagent's behavior.
- Ingestion points: The skill explicitly reads from
WebSearch,WebFetch, and MCP-connected tools (Notion, Confluence, Jira) as described inSKILL.md. - Boundary markers: The skill instructions recommend using "terse" requests and claim that untrusted content is isolated within child processes, which helps mitigate risks to the main agent.
- Capability inventory: The skill uses
nodeto execute local logic and utilizes subagents with read capabilities on the file system and external network resources. - Sanitization:
SKILL.mdstates that Web/MCP content is isolated in the child process to prevent injection attacks from affecting the parent agent. - [COMMAND_EXECUTION]: The skill utilizes a series of shell scripts as shims to invoke local Node.js logic.
- Evidence: Multiple files in the
scripts/directory, such asrun.sh,prepare.sh, anddispatch.sh, executenode delegate-cli.mjsto perform the skill's core functions. - [EXTERNAL_DOWNLOADS]: The
model-token-prices.jsonfile references documentation URLs from several well-known AI and technology service providers. - Evidence: The file includes reference URLs to official pricing pages for Anthropic, OpenAI, Cognition (Devin), Cursor, and OpenCode. These are well-known services and the references are documented neutrally for cost analysis purposes.
Audit Metadata