skills/zilliztech/mfs/open-tag-admin/Gen Agent Trust Hub

open-tag-admin

Pass

Audited by Gen Agent Trust Hub on Jun 25, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection.
  • Ingestion points: In scripts/slack_socket_agent.py, the skill retrieves the full Slack thread history using the conversations.replies API method.
  • Boundary markers: The retrieved thread content is interpolated into a system prompt in scripts/opentag_agent.py inside a block labeled 'Slack thread context', but without strong structural delimiters or instructions to ignore embedded commands.
  • Capability inventory: The skill can invoke high-autonomy agents (claude or codex) that have file system access and shell execution capabilities.
  • Sanitization: There is no sanitization or escaping of the user-provided thread content before it is passed to the AI backend.
  • [COMMAND_EXECUTION]: The skill uses subprocess.run to execute CLI agents.
  • In scripts/opentag_agent.py, the backends are invoked with flags that explicitly bypass security constraints, such as --dangerously-skip-permissions (for Claude) and --dangerously-bypass-approvals-and-sandbox (for Codex).
  • This grants the agent full autonomy over the host workspace and inherited environment based on input derived from untrusted Slack messages.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 25, 2026, 02:14 PM
Security Audit — agent-trust-hub — open-tag-admin