full-workflow
Pass
Audited by Gen Agent Trust Hub on Oct 17, 2026
Risk Level: SAFEPROMPT_INJECTIONPERSISTENCEINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill includes instructions in the 'Self-Evolving Skill' and 'Post-Execution Reflection' sections that command the agent to modify the SKILL.md file. This pattern encourages the agent to rewrite its own operational logic based on its experiences during execution, which could be exploited to inject and persist malicious instructions if the agent is manipulated into 'correcting' its behavior with attacker-controlled content.
- [PERSISTENCE]: By instructing the agent to 'fix this file immediately' and 'update this SKILL.md', the skill creates a mechanism for behavioral changes to persist across all future sessions. This allows any accidental or malicious modification of the instructions to become permanent until manually reverted.
- [INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves converting and analyzing terminal recordings, which are external data sources.
- Ingestion points: Reads terminal recording files (.cast) and processes outputs from the asciinema-analyzer tool.
- Boundary markers: None; the skill does not specify markers or instructions to help the agent distinguish between data to be analyzed and instructions embedded within that data.
- Capability inventory: The skill uses the Bash tool for command execution and the Write tool to modify files.
- Sanitization: There are no provisions for sanitizing terminal recordings or tool outputs, increasing the risk that embedded commands or adversarial text could influence the agent's behavior during the reporting or reflection phases.
Audit Metadata