self-debug
Warn
Audited by Gen Agent Trust Hub on Sep 8, 2026
Risk Level: MEDIUMPERSISTENCEMETADATA_POISONINGPROMPT_INJECTION
Full Analysis
- [PERSISTENCE]: The skill explicitly directs the agent in 'Layer 5: Solidify' to write discovered workaround methods and logic into its own 'SKILL.md', 'reference/', and 'MEMORY.md' files. This allows behavioral changes to persist across future sessions.
- [METADATA_POISONING]: The instructions encourage the agent to abstract rules from current tasks and insert them into 'SOUL.md' (defined as a cross-project general rule file). This allows for the injection of new instructions into the agent's governing logic based on runtime experiences which may be influenced by untrusted data.
- [PROMPT_INJECTION]: The skill contains 'overriding' instructions intended to bypass default agent behaviors, specifically telling the agent to 'go through the framework yourself first' and to avoid asking the user for help until all internal troubleshooting layers are exhausted.
- [INDIRECT_PROMPT_INJECTION]: By requiring the agent to 'spawn a sub-agent' and 'follow the action suggestions in the report' when stuck, the skill creates an execution chain where the primary agent's actions are governed by the output of a secondary process, which could be influenced if the troubleshooting context contains adversarial content.
Audit Metadata