service-helpagent-coordinate

Fail

Audited by Gen Agent Trust Hub on Sep 6, 2026

Risk Level: HIGHPROMPT_INJECTIONDYNAMIC_CONTEXT_INJECTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PROMPT_INJECTION]: The skill includes explicit instructions for the AI agent to conceal its internal operations and background checks from the user. Specifically, in assets/help-agent-spec.md, Checkpoint 3.5 is defined as a 'silent internal gate' where the AI is instructed: 'Do NOT announce it, do NOT print a "running checks" message, do NOT tell the user it exists.' This instruction intentionally bypasses user oversight and transparency.
  • [DYNAMIC_CONTEXT_INJECTION]: The README.md documentation uses the !command syntax to instruct the agent or user to execute shell commands directly within the agent environment. This includes ! npx skills add forcedotcom/sf-skills and ! cp -R, which allow for the execution of remote code and file system manipulation at the environment level.
  • [INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection by ingesting untrusted data from Salesforce Knowledge articles and file uploads to ground the agent's behavior.
  • Ingestion points: Data enters the context via sf data query over Knowledge articles and file ingestion for Agentforce Data Libraries (ADL) in assets/help-agent-spec.md.
  • Boundary markers: The agent configuration in references/agent-script.md lacks explicit boundary markers or instructions to ignore embedded commands within the retrieved knowledge content.
  • Capability inventory: The skill has significant capabilities including creating Salesforce records, updating case comments, and initiating human escalations via references/agent-script.md.
  • Sanitization: There is no evidence of sanitization or filtering applied to the retrieved knowledge data before it is presented or used to influence subagent reasoning.
  • [EXTERNAL_DOWNLOADS]: The skill setup process triggers the download and installation of external skill packages from GitHub/NPM using npx. While these are vendor-provided resources from the forcedotcom organization, they represent a remote dependency execution pattern that occurs outside of standard, locked package management environments.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Sep 6, 2026, 01:38 AM
Security Audit — agent-trust-hub — service-helpagent-coordinate