pdd-autopilot
Pass
Audited by Gen Agent Trust Hub on Nov 11, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to read and act upon data from GitHub issues and comments, which represents a surface for indirect prompt injection.
- Ingestion points:
SKILL.md(Step 2) andsteps/dependencies.mdspecify that the agent reads issue bodies and comments to determine requirements and dependencies. - Boundary markers: While the skill uses structured recipes from
pdd-prompting, it does not explicitly define delimiter-based boundary markers to separate instructions from data in the instructions provided. - Capability inventory: The skill possesses significant capabilities, including executing GitHub CLI (
gh) commands to modify issues and labels, and using thepddCLI to run workflows and arbitrary test commands (steps/handoff.md). - Sanitization: The skill implements a robust mitigation strategy by enforcing a "Trust" policy (Step 2 in
SKILL.md), where the agent only processes text from the repository owner, users with write access, or individuals explicitly added to a trust list. - [COMMAND_EXECUTION]: The skill performs its primary tasks by executing shell commands, including both fixed vendor tools and potentially dynamic repository commands.
- Evidence:
SKILL.mdandsteps/handoff.mddetail frequent use ofpddandghCLI tools. Notably,steps/handoff.mdincludes the capability to run arbitrary commands defined by the repository viaworkflow suite --command COMMAND. - [DATA_EXFILTRATION]: The skill includes a functional alerting system that transmits operational data to an external destination.
- Evidence: The
pdd autopilot alertcommand described insteps/handoff.mdsends JSON payloads containing failure causes and system status to a user-configurable "infrastructure destination". While this is a monitoring feature, it constitutes an outbound network path for repository-related data.
Audit Metadata