Security audit
Security checks for vulnerabilities and agentic risk
This skill is a disclosed every-turn orientation and local audit-log helper with no executable code, network access, credentials, or system-level authority.
Before installing, be comfortable with a skill that activates on every interaction and keeps a local audit log under ./memory/kernel/state.md. Review or clear that file if you do not want orientation outcomes retained across turns.
The skill is configured with an every_turn trigger, which causes it to activate on all interactions regardless of user intent. This expands the skill’s reach, increases the chance of unintended side effects such as repeated logging or policy interference, and makes any future unsafe behavior in the skill run far more often than necessary.
Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.
1\. \*\*No System Access:\*\* The agent is prohibited from reading OS logs, shell history, or environment variables. "Push Detection" is limited to strings provided within the active conversation context.
2\. \*\*Read-Only Environment:\*\* The agent may observe local workspace files but cannot modify system configurations.
3\. \*\*Mandatory Approval:\*\* Level 3 (Recursion) proposals are \*\*non-executable\*\*. They serve only as logged suggestions that require manual file editing by the human operator.
The skill instructs the agent to record outcomes to ./memory/kernel/state.md, which is a persistent file write, but this modification is not surfaced as a clear user-facing warning at the point of behavior. Silent or ambient persistence can surprise users, create privacy concerns, and lead to repeated accumulation of conversation-derived data across turns.
This plain-text skill file describes a mandatory behavior ('Before responding, perform this mental check') but does not define when the skill is invoked or what exact triggers activate it. Without explicit scope, trigger phrases, or exclusion conditions, the instructions are overly broad and could be applied to ordinary interactions unintentionally.
The skill is configured with an every_turn trigger, which causes it to run on every interaction rather than only in narrowly scoped contexts. This broad invocation increases the attack surface and the chance of unintended interference, prompt injection exposure, or policy-shaping behavior across unrelated conversations.
No suspicious patterns detected.