ask-pplx

Pass

Audited by Gen Agent Trust Hub on Nov 21, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes shell commands to interact with the local Ollama service and execute utility scripts for request validation.
  • Evidence: SKILL.md and references/ollama-api.md provide instructions for the agent to use ollama serve, ollama pull, curl, and python3 to manage the local environment and transport decision requests.
  • [EXTERNAL_DOWNLOADS]: The skill facilitates the download of model weights from the official Ollama model registry, which is a well-known service.
  • Evidence: README.md and references/ollama-api.md include commands such as ollama pull pplx-decider to fetch the necessary inference weights to the user's local machine.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes local repository content (code, configuration, and documentation) as evidence for model judgments, creating a surface for indirect prompt injection.
  • Ingestion points: SKILL.md Step 3 ("Gather the state") describes reading current code, configuration, screenshots, and specifications.
  • Boundary markers: The skill includes explicit instructions for the agent to treat retrieved material as evidence/data rather than instructions and provides a specific test case (injected-note in tests/cases.json) to verify the model's resilience.
  • Capability inventory: The skill is restricted to local shell calls and loopback network operations; it does not have autonomous write or external network capabilities beyond the local API.
  • Sanitization: Instructions in SKILL.md and references/authorization.md mandate the exclusion of credentials and personal data from the evidence state before transmission to the model.
  • [DYNAMIC_EXECUTION]: The package's test suite employs dynamic loading techniques to verify the integrity and contract of its internal scripts.
  • Evidence: tests/test_replay.py uses the importlib module to load and execute scripts/replay.py within the unit testing environment.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 21, 2026, 07:28 AM