The Agent Skills Directory

[COMMAND_EXECUTION]: Insecure deserialization pattern detected using the pickle library.
Evidence: The skill provides examples of using pickle.load(f) and pickle.dump() to manage document metadata associated with vector indices in references/vector-indexing.md and references/embedding-training.md.
Risk: If an attacker provides a malicious .docs file, loading it with pickle can lead to arbitrary code execution. This is a common pattern in machine learning utilities but remains a security concern when handling untrusted data.
[DATA_EXFILTRATION]: Functional network operations to well-known service providers.
Evidence: Integration with OpenAI and Cohere APIs for generating embeddings as documented in references/embedding-api-integration.md and references/embedding-models.md.
Context: These operations are consistent with the skill's primary purpose and use well-known, trusted domains (api.openai.com, api.cohere.com).
[PROMPT_INJECTION]: Indirect prompt injection vulnerability surface.
Ingestion points: The skill is designed to ingest and process large volumes of untrusted text data via functions like embed_pipeline, fixed_size_chunks, and semantic_chunks in SKILL.md and references/chunking-strategies.md.
Boundary markers: While the protocol suggests stripping HTML/Markdown, it lacks robust delimiter enforcement or instructions to the agent to ignore prompt-like content within the ingested documents.
Capability inventory: The skill possesses the ability to write to the local filesystem (indexing/caching) and make outbound network requests (API embedding services).
Sanitization: There is no evidence of sanitization to prevent adversarial instructions embedded in processed documents from influencing the agent's behavior during the embedding or retrieval phases.

ai-embeddings