extension-oql
Pass
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The ingestion tool (
scripts/ingest.mjs) executes theicpcommand-line utility to communicate with the Internet Computer. - Evidence: The
icpCallfunction (lines 186-210) usesnode:child_process.execFileto invokeicp canister callwith parameters for identity, network, and canister target. - Mitigation: The script validates canister identifiers using regex and escapes argument strings to mitigate injection risks.
- [INDIRECT_PROMPT_INJECTION]: The skill's function of exposing canister data to AI agents creates an indirect prompt injection attack surface.
- Ingestion points: Data is queried through the OQL
schema()andexecute()endpoints exposed by the canister. - Boundary markers: Data is returned to the agent in structured JSON format.
- Capability inventory: The agent can read all records defined as OQL entities by the canister developer.
- Sanitization: The skill does not implement specific sanitization of the data values retrieved from the canister.
- [EXTERNAL_DOWNLOADS]: The skill requires a Motoko package to be installed.
- Evidence:
SKILL.mdprovides instructions to runmops add caffeineai-oql@0.8.1. - Note: The package is a vendor-owned resource authored by caffeinelabs.
Audit Metadata