opportunity-factory
Pass
Audited by Gen Agent Trust Hub on Aug 13, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill includes utility and testing scripts that execute system commands.
- Evidence:
scripts/smoke_test_initializers.pyusessubprocess.runto executeinit_factory_workspace.pyandinit_factory_sqlite.pyduring validation. - Context: These are used for internal development, smoke testing, and workspace initialization within the skill's own environment.
- [INDIRECT_PROMPT_INJECTION]: The skill's primary function involves discovery tasks that ingest and analyze untrusted external data.
- Ingestion points:
discoverandresearchtasks (defined inreferences/workflow.md) mine data from external sources such as web search results, social media, app reviews, and forums. - Boundary markers: While the system uses structured JSON and Markdown, it lacks explicit "ignore instructions" delimiters when the worker role processes raw external text strings.
- Capability inventory: The agent has capabilities for file system writes (state and artifacts), subprocess execution (initialization scripts), and network operations (web fetching).
- Sanitization: No explicit sanitization or escaping of external content is documented before it is interpolated into local artifacts or state logs.
- [EXTERNAL_DOWNLOADS]: The skill facilitates automated data collection from external web sources.
- Evidence:
references/fallback-lane.mdandreferences/workflow.mddescribe roles for fetching web search results and API data for evidence collection. - Context: These are read-only operations intended for market research and are subject to the
autobucket approval policy, which forbids sending sensitive data externally. - [DYNAMIC_EXECUTION]: The skill includes a "Prompt Self-Improvement Loop" allowing it to modify its own internal instructions.
- Evidence:
references/prompt-self-improvement.mddetails how the agent can edit workspace-local prompts (.github/prompts/*.md). - Context: This autonomous modification is restricted to local workflow assets and requires a Layer 3 blocking critic (from a different model family) and a successful smoke test before changes are committed.
Audit Metadata