The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (MEDIUM): The skill repeatedly promotes the use of bun x repomix (in SKILL.md and references/architect-archive.md). This command downloads and executes the repomix package from a remote registry at runtime without version pinning, which is equivalent to unverified remote code execution.
[DATA_EXFILTRATION] (MEDIUM): The core 'Context Packing' strategy involves aggregating large segments of a codebase into a single file for ingestion by an external LLM. This practice encourages the bulk transfer of potentially proprietary source code to non-whitelisted third-party providers like OpenRouter.
[COMMAND_EXECUTION] (LOW): The skill provides examples of commands like git grep and rg for searching sensitive strings (e.g., SECRET, PASSWORD). While standard for developers, an autonomous agent using these tools could be manipulated into exposing credentials stored in the repository.
[REMOTE_CODE_EXECUTION] (HIGH): The script scripts/generate_image.py presents an Indirect Prompt Injection surface. If an agent populates the prompt or processes an input_image from untrusted sources, it could be coerced into performing unintended actions through the model interface.
Ingestion points: prompt argument and --input file path in scripts/generate_image.py.
Boundary markers: Absent; there are no delimiters or instructions to isolate the user-provided prompt from the agent's instructions.
Capability inventory: The script performs network requests (via requests.post) to an external API and writes data to the local filesystem (open(output_path, 'wb')).
Sanitization: Absent; input strings are passed directly to the external API without validation or escaping.

ai-pro