triple-loop-learning
Pass
Audited by Gen Agent Trust Hub on Jun 12, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill implements an architecture that ingests external data to drive its self-improvement logic, creating an indirect prompt injection risk.
- Ingestion points: SKILL.md (Step 1): "ingests execution logs from existing operations".
- Boundary markers: None identified; there are no instructions to the agent to treat log data as untrusted or to ignore embedded commands.
- Capability inventory: The skill has access to
Bash,Read, andWritetools. It is designed to perform "Code Mutation" and run "headless deterministic tests" (Step 4). - Sanitization: None identified; the skill does not describe any validation or filtering of the ingested log content.
- [COMMAND_EXECUTION]: The workflow involves the autonomous generation and execution of code. In Step 4, the agent is instructed to "Apply the instruction set or code adjustment" and "Run pure, headless deterministic tests" using the
Bashtool. This cycle of generating and executing content poses a risk of executing malicious logic if the mutation phase is influenced by adversarial input.
Audit Metadata