jflow
Pass
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes shell commands specified in the 'Check' column of the Proof table within '.jflow//spec.md' during the 'prove' and 'done' steps. It also runs security auditing tools like 'gitleaks' and various repository-specific verification commands.\n- [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection via the specification file. The agent executes commands found in the Proof table, which are derived from potentially attacker-influenced sources such as issue descriptions and codebase content. If malicious instructions are incorporated into the spec, they would be executed with the agent's privileges.\n
- Ingestion points: '.jflow//spec.md' (generated from issues and code, then read by multiple agent roles).\n
- Boundary markers: No delimiters or safety instructions are used to distinguish executable commands from descriptive text within the spec file.\n
- Capability inventory: Shell command execution, git repository operations, file system access, and model memory modification.\n
- Sanitization: There is no validation or sanitization of commands extracted from the specification file prior to execution.\n- [DYNAMIC_EXECUTION]: During the 'prove' step, the agent is directed to write and run 'driver' scripts for automated tests that do not yet exist. This involves runtime code generation and execution by the agent.\n- [COMMAND_EXECUTION]: The main 'SKILL.md' file recommends enabling network access in the execution sandbox (specifically via 'sandbox_workspace_write.network_access=true' for Codex). While intended for testing, this configuration could be abused to facilitate data exfiltration if the agent is manipulated into running malicious commands.
Audit Metadata