scoped-grill-with-docs

Warn

Audited by Gen Agent Trust Hub on Aug 7, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructs the agent to 'Resolve the operating system's temporary directory with a platform-native facility' and 'Create a uniquely named Markdown file'. This requires the agent to probe the host environment and perform file-system operations using shell commands or system APIs.
  • [PROMPT_INJECTION]: The skill implements a state-management mechanism called a 'scope ledger' that is susceptible to indirect prompt injection. User answers are written to a file and subsequently 'reread' by the agent to influence its decision-making logic ('Compare the complete proposed question with IN SCOPE'). There are no boundary markers or instructions to sanitize or ignore potential commands retrieved from the ledger content.
  • [DATA_EXFILTRATION]: By instructing the agent to write session data ('TL;DR' and 'IN SCOPE' items) to the local file system, the skill exposes internal agent state to the host environment. This represents a potential data exposure risk if the environment is not adequately isolated.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Aug 7, 2026, 06:00 PM
Security Audit — agent-trust-hub — scoped-grill-with-docs