The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the user's screen, including window titles and OCR text, which is subsequently processed by an LLM to generate code and skill definitions. This creates an attack surface where malicious content on the user's screen could influence the generation of skill drafts.
Ingestion points: scripts/fetch_window.py retrieves window titles and text content from the local Screenpipe API.
Boundary markers: The synthesis prompt in scripts/synthesize.py lacks clear delimiters or protective instructions to distinguish observed user data from system commands.
Capability inventory: The skill possesses the ability to write files to the local disk (scripts/run.py), move directories between folders (scripts/promote.py), and perform network operations to LLM backends.
Sanitization: The skill employs regex-based redaction via scripts/redact.py to strip PII and secrets, but it does not include filters or sanitization for prompt injection or command-like sequences.
[DATA_EXFILTRATION]: The skill processes highly sensitive screen capture history and is capable of transmitting derived summaries to external LLM providers.
Evidence: scripts/backends.py implements logic to send cluster summaries to Anthropic Claude or other user-defined cloud endpoints if the user opts out of the local-only default.
Privacy Controls: Mitigation measures include local processing as the default, mandatory PII redaction before transmission, and security checks in scripts/backends.py to ensure that API keys and sensitive data are not sent over plaintext HTTP to remote hosts.
[COMMAND_EXECUTION]: The skill automates file system management tasks to handle skill proposals.
Evidence: The scripts/promote.py script utilizes shutil.move to relocate directories, and scripts/run.py writes new SKILL.md markdown files to the local filesystem.
[EXTERNAL_DOWNLOADS]: The skill downloads an external machine learning model to facilitate semantic matching of user workflows.
Evidence: scripts/run.py uses the sentence-transformers library to fetch the all-MiniLM-L6-v2 model from HuggingFace.

autoskill