The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches and processes social media posts, profile descriptions, and interaction data from the ATProto network. This content originates from external users and is considered untrusted, creating a potential surface for indirect prompt injection where malicious instructions could be embedded in post text or profile data.
Ingestion points: Data is ingested through multiple functions in atproto.py (e.g., records, interactions, thread) and feed.py (e.g., collect, fetch_post) which query external servers for network records.
Boundary markers: The processed data is returned to the agent as text digests. There are no explicit delimiters or warning instructions provided to the agent to differentiate this untrusted content from the skill's own instructions.
Capability inventory: The skill possesses the ability to make network requests to external domains and write HTML files to the local file system using the --html argument in the feed command.
Sanitization: While the skill performs basic cleaning of whitespace for display purposes, it does not implement specific security sanitization or filtering to detect or neutralize prompt injection attempts within social media content.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to retrieve data from decentralized ATProto infrastructure.
Evidence: The http function in scripts/atproto.py makes requests to plc.directory for identity resolution, constellation.microcosm.blue for backlink indexing, ufos-api.microcosm.blue for lexicon activity monitoring, and various user-specific Personal Data Servers (PDS) to fetch authoritative repo records.
[COMMAND_EXECUTION]: The skill is designed to be operated via a command-line interface provided to the AI agent.
Evidence: The SKILL.md and README.md files provide instructions for the agent to execute python3 scripts/atproto.py with various subcommands to perform queries and reconstruct timelines.

atprotoing