fleet
Pass
Audited by Gen Agent Trust Hub on Nov 18, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill provides functionality to import GitHub issues (via
fleet import) which are then processed as tickets by AI agents. These tickets can define shell commands for testing and verification (via thefleet gatecommand), creating a potential surface for an attacker to inject malicious commands into an issue body. - Ingestion points: GitHub issue titles and bodies are imported into local markdown files within the
.scratch/directory. - Boundary markers: The instructions mandate the use of extended code fences and explicit warnings that the issue body is untrusted data and should not be treated as instructions.
- Capability inventory: The system is capable of arbitrary shell command execution through the
fleet gatetool and can manage multiple interactive CLI environments. - Sanitization: The skill relies on the lead agent's ability to distinguish data from instructions and requires owner approval of the planned execution graph before any commands are run.
- [DYNAMIC_EXECUTION]: The helper script
bin/fleet-statuslineutilizes Python'sexec()function to load its main logic from a local file namedfleetat runtime. While this is used to implement custom bytecode caching, it involves dynamic execution of code determined by the script's directory path. - [COMMAND_EXECUTION]: The
bin/fleet-statuslinescript usessubprocess.Popenwithshell=Trueto execute a command provided by the user via the--thenargument. This is intended for chaining existing terminal status lines but represents a sensitive execution pattern if the input is influenced by an attacker.
Audit Metadata