zcode-delegate
Warn
Audited by Gen Agent Trust Hub on Sep 2, 2026
Risk Level: MEDIUMCREDENTIALS_UNSAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONREMOTE_CODE_EXECUTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill resolves the path to the ZCode CLI at runtime using environment variables or command-line arguments. This allows for the dynamic loading of executable content from computed paths, which could be exploited to point the agent to malicious binaries.
- Evidence: References to resolving the CLI via the
ZCODE_CLIenvironment variable or the--zcode-pathflag. - [DATA_EXFILTRATION]: The skill documentation encourages interactions with sensitive local configuration files that store authentication data and provider settings.
- Evidence: Instructions for reading and writing to
~/.zcode/cli/config.jsonto configure model providers. - [INDIRECT_PROMPT_INJECTION]: The skill creates a multi-agent vulnerability surface by ingesting untrusted task descriptions (briefs) and relaying them to an external autonomous agent (ZCode) that possesses broad system capabilities.
- Ingestion points: Task briefs are accepted via the
--briefflag (file-based) or through standard input. - Boundary markers: The brief is passed to the ZCode CLI as a file attachment using the
--attachflag to separate it from the initial prompt. - Capability inventory: The sub-agent (ZCode) is capable of writing files and executing shell commands when running in
yolomode. - Sanitization: There is no evidence of sanitization, filtering, or validation of the brief's content before it is delivered to the sub-agent.
- [COMMAND_EXECUTION]: The skill mechanism relies on the execution of shell commands to manage the delegation loop and interact with the ZCode CLI.
- Evidence: Implementation via
node scripts/relay.mjswhich spawns and manages child processes. - [CREDENTIALS_UNSAFE]: The skill requires the user to provide sensitive authentication secrets to facilitate the delegation process.
- Evidence: Instructions to set and use environment variables such as
ZAI_API_KEY,ZCODE_API_KEY, andANTHROPIC_API_KEYto authenticate the headless CLI. - [REMOTE_CODE_EXECUTION]: The skill depends on an external, unverifiable CLI tool that is not distributed through standard, trusted package registries and is resolved from arbitrary local application paths.
- Evidence: The
zcodeCLI is described as shipping inside a desktop app bundle rather than being available vianpmor standard system package managers.
Audit Metadata