issue-create
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONMETADATA_POISONINGCOMMAND_EXECUTIONCREDENTIALS_UNSAFE
Full Analysis
- [METADATA_POISONING]: The skill includes a 'Self-Evolving Skill' directive in SKILL.md that instructs the agent to 'fix this file immediately, don't defer' if issues are encountered. This encourages autonomous self-modification of the skill's own instructions, which could be exploited to remove safety constraints or introduce malicious behavior.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data (issue bodies) and repository metadata (labels) and interpolates them directly into prompts for AI classification and labeling.
- Ingestion points: Untrusted data enters the agent context via the
--bodyCLI argument and repository label lists fetched from theghCLI. - Boundary markers: The prompt templates in
references/ai-prompts.mdlack delimiters or explicit instructions to ignore embedded commands within the{content}and{body}placeholders. - Capability inventory: The skill has
Bashaccess and utilizes theghandgitCLIs, which allows it to read/write repository data, create issues, and access local logs and caches. - Sanitization: There is no evidence of sanitization, escaping, or filtering of the issue content before it is processed by the AI models.
- [COMMAND_EXECUTION]: The skill performs various shell operations using
Bash, theghCLI, andgitto manage repository context, fetch taxonomy, and submit issue data. - [CREDENTIALS_UNSAFE]: The documentation describes the use of a persistent Playwright browser profile (
~/.claude/tools/pw-github-profile/) for automated image uploads. This profile is a sensitive directory that typically contains authenticated session cookies and credentials for GitHub access.
Audit Metadata