Security audit
Security checks for vulnerabilities and agentic risk
This is a markdown-only agent failure-handling skill with no executable code, network behavior, or hidden install actions.
Installers should be aware that this skill may activate during ordinary uncertainty or correction and may write small preference notes if the host agent has durable memory enabled. That behavior is disclosed and narrow, but users who do not want persistent working-style notes should disable memory or adapt that section before use.
The README says to use the skill when the agent sees conditions like 'low confidence,' 'plan collapse,' or 'harsh or direct correction,' but these terms are subjective and not bounded by explicit thresholds or exclusions. Because the trigger scope is not narrowly defined and includes common runtime situations, the skill may activate more often than intended.
The skill instructs automatic activation on vague conditions like 'low confidence,' 'owner frustration,' or 'repeated failure' without clear thresholds or authorization boundaries. In an agent system, this can cause the skill to take over too broadly, altering behavior unexpectedly, suppressing normal task flow, or steering the agent into fallback mode when it should continue with the original task.
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
## Anti-Patterns
- Claiming success without verification.
- Apologizing repeatedly instead of changing behavior.
- Asking vague questions like "What would you like me to do?"
- Retrying with the same plan after the same failure.