autonomous-agent-patterns
Warn
Audited by Gen Agent Trust Hub on Apr 15, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill includes code templates for executing shell commands. Specifically, the 'SandboxedExecution' pattern utilizes subprocess.run with shell=True, which allows for shell interpolation and increases the risk of command injection.
- [REMOTE_CODE_EXECUTION]: The 'MCPAgent' architectural pattern demonstrates a method for the agent to generate code via an LLM and write it to a local file ('server.py') for execution. This facilitates the execution of dynamically generated, untrusted code.
- [DATA_EXFILTRATION]: The skill describes patterns for tools that can read local filesystem data ('read_file') and fetch content from remote URLs ('add_url' via the requests library), which are primary components of data exposure and exfiltration workflows.
- [PROMPT_INJECTION]: The skill describes surface areas for indirect prompt injection where the agent ingests data from untrusted sources. Evidence chain: 1. Ingestion points: ContextManager.add_url and BrowserTool.get_page_content. 2. Boundary markers: Absent in implementation patterns. 3. Capability inventory: run_command, write_file, and edit_file. 4. Sanitization: Absent in the provided snippets.
Audit Metadata