delegate-grok

Pass

Audited by Gen Agent Trust Hub on Aug 20, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it constructs instructions for a secondary AI (Grok) by interpolating untrusted task text directly into a command string.
  • Ingestion points: Data enters through the --task flag or standard input in the scripts/delegate-grok.sh script, which is then passed to the Grok agent via herdr.
  • Boundary markers: The script uses simple text labels such as 'Task:' and 'Rules:' to separate system instructions from the delegated content, which is insufficient to prevent an adversary from overriding the prompt context.
  • Capability inventory: The script utilizes the herdr CLI to start agents and send prompts, meaning any successfully injected instruction could potentially leverage the capabilities of the Grok agent within the shared Herdr workspace.
  • Sanitization: There is no escaping, filtering, or structured schema validation of the input task content before it is incorporated into the final prompt.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 20, 2026, 01:24 AM
Security Audit — agent-trust-hub — delegate-grok