canhelp
Pass
Audited by Gen Agent Trust Hub on Nov 19, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill fetches and summarizes Candid (.did) files from external canisters on the Internet Computer mainnet. These files are untrusted and could contain malicious documentation strings or method names designed to inject instructions into the AI during the summarization process. * Ingestion points: The file is downloaded to /tmp by
scripts/fetch-candid.shand read into context by theReadtool. * Boundary markers: There are no explicit delimiters or system instructions provided to the model to ignore instructions embedded within the Candid file. * Capability inventory: The skill usesBash(for local helper scripts),Read,Grep, andGlobtools. * Sanitization: The skill performs no filtering or sanitization of the Candid content before presenting it to the agent. - [EXTERNAL_DOWNLOADS]: The
scripts/resolve-canister-id.shscript makes network calls to the official Internet Computer Dashboard API (ic-api.internetcomputer.org) usingcurlto resolve human-readable canister names. - [COMMAND_EXECUTION]: The skill executes local bash scripts to interact with the
icpCLI and perform network lookups.
Audit Metadata