The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes the Turkish instruction "Kanitsal Cerceve (Evidential Frame Activation) Kaynak dogrulama modu etkin" (Evidential Frame Activation: Source verification mode active). This is a recognized prompt engineering technique intended to alter the AI's internal state or bypass behavioral constraints by imposing a specific linguistic "frame."
[DATA_EXPOSURE]: The skill instructs the agent to scan and extract content from D:\Projects\*. This represents a broad and potentially sensitive directory scan that could expose private source code, environment variables, or other sensitive project data.
[COMMAND_EXECUTION]: The skill uses bash -lc to run the codex binary while interpolating user-supplied experiment descriptions. This pattern creates a risk of command injection if the input strings are not properly sanitized.
[METADATA_POISONING]: There is a mismatch between the author listed in the SKILL.md metadata (context-cascade) and the reported vendor identity, which can lead to confusion regarding the skill's origin and safety.
[INDIRECT_PROMPT_INJECTION]: The skill has a vulnerable surface for indirect prompt injection.
Ingestion points: Reads files and logic from D:\Projects\* and .claude/library/catalog.json (SKILL.md).
Boundary markers: None. There are no instructions to disregard embedded commands in the files it scans.
Capability inventory: The skill uses Bash, Write, and Edit tools to execute code and modify the filesystem.
Sanitization: No sanitization logic is described for handling external file content before it is processed or adapted.

codex-safe-experiment