The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The SentenceTransformersMultimodalBackend class in scripts/core/embeddings/sentence_transformers_multimodal.py loads models using the trust_remote_code=True parameter. This setting allows the execution of arbitrary Python code defined within a remote model's repository (e.g., from Hugging Face) upon loading, which can lead to a full system compromise if a malicious model is selected.
[COMMAND_EXECUTION]: The skill makes frequent use of subprocess.run to interact with the host system. It invokes external binaries like textutil, antiword, and catdoc to parse documents, and executes a helper script word_com.py for Windows Word COM interaction. While these are intended functions, they expand the attack surface for command injection if file paths are manipulated.
[EXTERNAL_DOWNLOADS]: The local_knowledge.sh wrapper and the Python backend automate the download and installation of numerous third-party dependencies from PyPI and embedding models from Hugging Face. This occurs during the doctor --install phase and automatically if dependencies are missing during indexing.
[DATA_EXFILTRATION]: The skill is designed to recursively read and index local directories. While no explicit malicious exfiltration was found, the core functionality provides the agent with the ability to read and process sensitive local files, which could be leveraged if the agent's behavior is influenced by an attacker.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection by indexing untrusted document content.
Ingestion points: The scan_documents function in scripts/core/files.py reads content from various local file types.
Boundary markers: The skill lacks explicit boundary delimiters or safety instructions when processing or returning retrieved document chunks.
Capability inventory: The skill has significant capabilities, including file system access and shell command execution.
Sanitization: No sanitization or content filtering is performed on the data ingested from files.

local-knowledge