The Agent Skills Directory

[PROMPT_INJECTION] (HIGH): The skill employs 'Kanitsal Cerceve (Evidential Frame Activation)' and 'Kaynak dogrulama modu etkin' in SKILL.md. This is a non-standard framing technique, using foreign languages and specialized terminology to force the model into a specific state, often used to bypass safety filters.
[INDIRECT_PROMPT_INJECTION] (HIGH):
Ingestion points: The skill reads the entire codebase using gemini --all-files (SKILL.md, README.md).
Boundary markers: No boundary markers or 'ignore embedded instructions' warnings are present to isolate the untrusted codebase content.
Capability inventory: The skill allows high-privilege tools including Bash, Write, and Edit (SKILL.md).
Sanitization: No sanitization is performed. Malicious instructions inside the codebase could influence the agent's logic during the subsequent 'IMPLEMENT' phase, leading to unauthorized actions.
[COMMAND_EXECUTION] (HIGH): The skill uses bash -lc to execute commands that include user-provided query strings and external file content. The 'gemini-yolo.sh' script suggests a 'YOLO' (You Only Look Once/Live Once) approach, which often implies bypassing safety confirmations or checks.
[METADATA_POISONING] (MEDIUM): The x-verix-description field uses pseudo-technical assertive language ([assert|neutral], [state:confirmed]) to trick automated analyzers or the model into assuming the skill is verified and safe, violating the principle of authoritative self-claims.

gemini-codebase-onboard