The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes text from files on the local filesystem to generate embeddings. This creates an attack surface where malicious instructions embedded in those files could be retrieved and processed by the AI agent, potentially influencing its subsequent actions.
Ingestion points: scripts/semantic_grep.py loads file content via the load_corpus() function.
Boundary markers: No specific delimiters or safety warnings are added to the search results before returning them to the agent context.
Capability inventory: The skill performs network operations to external embedding APIs using requests.post and accesses the local filesystem.
Sanitization: None. File content is read directly and passed to the embedding provider.- [DATA_EXFILTRATION]: The skill reads local file content and transmits it to external endpoints for embedding generation. The load_corpus function in scripts/semantic_grep.py follows symbolic links when scanning directories, which could be exploited to read and transmit sensitive files outside the intended search path if a malicious symlink is present. The destination endpoints are well-known technology services.- [EXTERNAL_DOWNLOADS]: The skill communicates with Cloudflare AI Gateway (gateway.ai.cloudflare.com) and Google AI Studio (generativelanguage.googleapis.com) to generate embeddings. These are well-known services used for the skill's primary functionality.

semantic-grep