pdd-prompting
Pass
Audited by Gen Agent Trust Hub on Nov 21, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill's core workflow requires the agent to read and follow instructions contained in
.prompt.mdanduser_stories/*.mdfiles within the repository. This creates a surface for indirect prompt injection if an attacker can commit malicious content to these files. - Ingestion points: The agent is instructed to ingest data from files such as
prompts/*.prompt.md,user_stories/*.md, andtests/*.jsonas described inSKILL.md,RUNBOOK.md, andguide.md. - Boundary markers: The skill documentation mentions the use of
pdd-boundaryblocks to narrow scope (referenced inrunbook/files-and-runtime-block.md). - Capability inventory: The agent has the capability to execute shell commands (
pdd,gh), read and write to the filesystem, and interact with GitHub Pull Requests. - Sanitization: The workflow includes mandatory human approval gates (using
pdd approve --said) before code changes or merges can proceed, acting as a defense against unintended instruction following. - [COMMAND_EXECUTION]: The skill relies on the execution of local command-line tools to perform its functions.
- Evidence: Instructions in
SKILL.md,RUNBOOK.md, and other runbook files direct the agent to usepdd test,pdd lint,pdd simulate,gh pr merge, and other CLI-based tools to verify behavior and manage the development process. The skill explicitly warns against shell-interpolating untrusted words when commenting on PRs.
Audit Metadata