project-foundation

Warn

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill utilizes a 'curated pack' of external extensions (skills) that are downloaded from the internet during the intake or UX/UI phases. While some sources like Vercel Labs are well-known, others come from individual third-party accounts.
  • Evidence: npx -y skills add leonxlnx/taste-skill@design-taste-frontend -g -y in SKILL.md and assets/install-curated-pack.sh.
  • Evidence: npx -y skills add pbakaus/impeccable@impeccable -g -y in SKILL.md and assets/install-curated-pack.sh.
  • Evidence: npx -y skills add mattpocock/skills@grill-me -g -y in SKILL.md and assets/install-curated-pack.sh.
  • Evidence: npx -y skills add mattpocock/skills@improve-codebase-architecture -g -y in SKILL.md and assets/install-curated-pack.sh.
  • [REMOTE_CODE_EXECUTION]: The skill performs automated "Just-In-Time" (JIT) installation of remote code. If design-related skills are not detected during Phase 06, the orchestrator is instructed to install them automatically without further confirmation.
  • Evidence: If no skill is found, the orchestrator automatically installs design-taste-frontend with: npx -y skills add leonxlnx/taste-skill@design-taste-frontend -g -y and continues without stopping the flow. in phases/06-ux-ui.md.
  • [INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection by fetching and processing untrusted data from the public web.
  • Ingestion points: The research phase (phases/01-research.md) uses subagents (research-competitors.md, research-reviews.md) to fetch information about competitors and user reviews via WebSearch and WebFetch.
  • Boundary markers: The skill attempts to isolate this data in docs/research/notes/ and requires subagents to return compact YAML handoffs, which reduces but does not eliminate the risk.
  • Capability inventory: The orchestrator has extensive tools including Bash, Edit, Write, and Task, which could be exploited if the AI obeys instructions embedded in the fetched web content.
  • Sanitization: There is no evidence of content sanitization or filtering applied to the fetched web data before it is processed by the synthesis subagents.
  • [COMMAND_EXECUTION]: The skill executes local shell and PowerShell scripts provided within the skill package to manage state and install dependencies.
  • Evidence: Execution of assets/state-tool.sh, assets/state-tool.ps1, assets/install-curated-pack.sh, and assets/install-curated-pack.ps1 via the Bash tool.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 17, 2026, 02:31 AM