synesthesia
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines a workflow that ingests untrusted technical data to generate metaphorical mappings which are then persisted in a memory tool.
- Ingestion points: The skill reads external content such as code, logs, and architecture documentation during the 'Literal read' phase in
SKILL.md. - Boundary markers: The 'Core Contract' and 'Guardrails' sections provide instructions to the agent to prioritize literal evidence and avoid treating metaphors as proof.
- Capability inventory: The skill uses
run_memory_note_toolto write data to a durable memory store. - Sanitization: The 'Mapping Admission Gate' enforces a structured schema and requires explicit user endorsement before any data is written to the memory tool, acting as a manual verification layer.
- [COMMAND_EXECUTION]: The skill uses a command-line utility to manage its internal state.
- Evidence: The instruction
run_memory_note_tool append --extension synesthesia --kind mapping-endorsement --json -inSKILL.mdexecutes a shell command to persist JSON payloads.
Audit Metadata