mirofish-offline-simulation
Warn
Audited by Gen Agent Trust Hub on Sep 12, 2026
Risk Level: MEDIUMEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
- [METADATA_POISONING]: There is a discrepancy between the skill's declared author in the body ('ara.so') and the author context provided, which may be misleading regarding the source of the skill.
- [EXTERNAL_DOWNLOADS]: The skill provides instructions to clone the
nikmcfly/MiroFish-Offlinerepository. As this repository is not owned by a trusted organization, it requires manual review before execution. - [COMMAND_EXECUTION]: The installation instructions involve executing shell commands that download and run code, such as
docker compose,pip install, andnpm install. These actions execute code defined in the external repository. - [INDIRECT_PROMPT_INJECTION]: The skill processes external, potentially untrusted documents through an LLM to build its simulation environment.
- Ingestion points: Untrusted document content is read in
SKILL.mdand passed toGraphBuilder.build(). - Boundary markers: There are no explicit markers or instructions displayed in the snippets to prevent the LLM from following commands embedded in the documents.
- Capability inventory: The ingested data influences agent behavior and memory stored in Neo4j. The skill also invokes local LLM services via Ollama.
- Sanitization: No sanitization or validation of the input text is mentioned before it is processed by the graph builder.
Audit Metadata