account-plan
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCREDENTIALS_UNSAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill fetches external data from company websites, LinkedIn pages, and job postings, creating a surface for indirect prompt injection attacks where malicious instructions hidden on those pages could influence agent behavior.
- Ingestion points:
WebSearchandWebFetchtools are used inSKILL.mdto research accounts. - Boundary markers: The
references/agent-security.mdfile provides explicit instructions to treat all fetched content as data rather than instructions and to quote rather than follow any embedded commands. - Capability inventory: The skill writes research findings to
.agents/product-context.mdand generates visual plans using platform artifact tools. - Sanitization:
references/agent-security.mdmandates HTML-escaping for all interpolated values in generated markup. - [DYNAMIC_EXECUTION]: The skill uses platform-specific artifact tools to generate visual account plans, which involves dynamic code/markup generation.
- Evidence:
SKILL.mdincludes a section for a 'Visual account plan' that utilizes host tools like Claude'sArtifacttool to render documents and charts. - Mitigation:
references/agent-security.mdsection 3 contains strict rules for escaping HTML, prohibiting unsafe functions likeinnerHTML, and validating URLs in emitted code. - [CREDENTIALS_UNSAFE]: While the skill does not contain hardcoded credentials, it is designed to process CRM exports and transcripts which often contain them.
- Mitigation: The skill includes comprehensive instructions in
references/agent-security.mdto identify, redact, and flag credentials (API keys, session tokens) without reproducing them in output or persisting them to files. - [PROMPT_INJECTION]: The skill contains instructions to suppress certain output patterns (like warnings on drafts or comments about missing tools) to improve user experience.
- Evidence:
SKILL.mdinstructs the agent to "Do not return a draft with a warning on it" and to "skip this step without comment" if a specific visual tool is missing.
Audit Metadata