The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements a 'Self-Evolution Subsystem' (detailed in references/self-evolution.md) that enables the agent to modify its own instruction files and detection patterns. Under 'Safety Level B', the agent can autonomously refine criteria and patterns in the references/ directory without human review, provided the changes pass a self-audit. This allows the agent to dynamically alter its behavioral logic and audit standards at runtime.
[PERSISTENCE]: The self-evolution mechanism functions as a persistence vector by writing updates to the agent's filesystem-based instruction files. Modifications made during the evolution cycle persist across different sessions, allowing the agent's logic and ruleset to be permanently altered by external data or user interactions processed during the 'RESEARCH' and 'UPDATE' phases.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection. It ingests external SKILL.md files for auditing and processes 'user corrections' (trigger GT-03) to inform its evolution. Malicious content in these inputs could potentially poison the agent's detection patterns or checklist criteria, particularly when changes are applied autonomously under 'Safety Level B'.
Ingestion points: External SKILL.md files being audited and 'user corrections' provided during the feedback phase.
Boundary markers: No specific delimiters or 'ignore' instructions are defined for processing user-provided logic corrections.
Capability inventory: Ability to perform web searches and execute file-write operations to the references/ directory.
Sanitization: The skill does not explicitly describe sanitization or validation protocols for user-provided data incorporated into the evolution cycle.
[TIME_DELAYED_ATTACKS]: The agent includes time-delayed trigger logic (GT-04) that initiates a full research and self-evolution cycle after 30 days of inactivity. This conditional behavior gates network operations and file modifications based on an elapsed time threshold rather than direct user intent.

gauge