Livestreaming to ATProtocol

Warn

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: MEDIUMDATA_EXFILTRATIONCOMMAND_EXECUTIONCREDENTIALS_UNSAFE
Full Analysis
  • DATA_EXFILTRATION (MEDIUM): The skill transmits agent internal state, including 'reasoning_message' and 'assistant_message' content, to an external PDS (Personal Data Server) on the ATProtocol network.
  • Evidence: httpx.post calls to {PDS}/xrpc/com.atproto.repo.createRecord in references/livestream.py and references/publish-response.py.
  • Risk: Internal 'reasoning' often contains raw data processed by the agent that regex-based redaction (defined in REDACT_PATTERNS) may not capture, leading to potential exposure of private information to a public or semi-public feed.
  • COMMAND_EXECUTION (LOW): The skill's documentation instructs the user to configure system hooks that execute Python scripts via shell commands.
  • Evidence: SKILL.md setup section includes shell commands for hook execution: uv run python hooks/livestream.py.
  • CREDENTIALS_UNSAFE (LOW): The scripts use a hardcoded absolute path to a .env file for loading sensitive ATProtocol and API credentials.
  • Evidence: load_dotenv("/home/cameron/central/.env") found in both reference scripts.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Feb 17, 2026, 06:49 PM