The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides detailed patterns and instructions for spawning child processes and executing shell commands through the ChildProcess and ChildProcessSpawner services.
Evidence: Examples such as ChildProcess.make('ls', ['-la']) and complex command pipelines using ChildProcess.pipeTo for cat, grep, and wc.
[CREDENTIALS_UNSAFE]: A code snippet in the documentation illustrating how to configure environment variables for a child process includes a hardcoded secret value.
Evidence: The configuration { env: { NODE_ENV: 'production', API_KEY: 'secret' } } found in the Advanced ChildProcess Usage section of SKILL.md.
[DATA_EXFILTRATION]: The skill facilitates the reading of local filesystem data and its subsequent transmission over the network, establishing a functional path for data exfiltration.
Evidence: The instructions demonstrate using FileSystem.readFileString to ingest local data and HttpClient.post to send data to external endpoints.
[EXTERNAL_DOWNLOADS]: The skill documents the implementation of outbound network requests to external APIs using the HttpClient service.
Evidence: Code examples show GET and POST requests to https://api.example.com/ endpoints.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests data from untrusted sources (files, user input, and network responses) and provides access to high-privilege services without demonstrated sanitization.
Ingestion points: Data entering through FileSystem.readFileString, Terminal.readLine, and HttpClient response parsing.
Boundary markers: No delimiters or protective instructions are included in the documented patterns to prevent the agent from executing instructions embedded in processed data.
Capability inventory: Includes services for process execution (ChildProcessSpawner), filesystem writes (FileSystem), and network transmission (HttpClient).
Sanitization: No evidence of data validation, escaping, or sanitization logic is present in the provided examples.

effect-platform-abstraction