postplus-shared

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill contains instructions to omit technical internals and implementation details from user-facing communication, which can reduce transparency.
  • Evidence: references/shared-user-guidance.md directs the agent to 'Avoid unnecessary provider, API, schema, endpoint, payload, base64, runtime, attachment handle, or storage internals' and to 'Keep detailed technical evidence in local artifacts, logs, or task summaries unless the user asks for implementation details.'
  • [COMMAND_EXECUTION]: The agent is instructed to execute specific command-line operations for maintenance and session management using vendor-owned tools.
  • Evidence: references/shared-public-skill-rules.md instructs the agent to run postplus update, postplus auth login, and postplus quote confirm commands.
  • [EXTERNAL_DOWNLOADS]: The skill directs the agent to proactively install missing third-party dependencies in the local environment.
  • Evidence: references/shared-public-skill-rules.md identifies python3, yt_dlp, ffmpeg, and ffprobe as approved dependencies and instructs the agent to 'Use the smallest direct install path already supported by the host environment' to perform the installation.
  • [INDIRECT_PROMPT_INJECTION]: The guidelines define a high attack surface for indirect prompt injection by documenting the ingestion of large amounts of untrusted data from multiple social platforms.
  • Ingestion points: references/dataset-item-schemas.md lists extensive data fields retrieved from Facebook, Instagram, TikTok, Reddit, X (Twitter), and YouTube.
  • Boundary markers: references/research-quality-recovery.md requires record inspection for quality but lacks specific instructions to ignore embedded commands or use boundary delimiters.
  • Capability inventory: references/shared-public-skill-rules.md documents capabilities for subprocess execution (via vendor CLI and dependency installers) and local file system writes.
  • Sanitization: No explicit sanitization or content filtering rules are specified for the external data being processed.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 12:59 AM