The Agent Skills Directory

[COMMAND_EXECUTION]: The code templates (templates/train_template.py and templates/vllm_adapter_template.py) include a copy_file utility function using shutil.copy2. This function is exposed via command-line arguments (--copy_from and --copy_to), allowing the generated scripts to move arbitrary files on the local filesystem if executed by the agent or a user.
[REMOTE_CODE_EXECUTION]: The skill configuration and documentation (e.g., references/vllm_integration.md and references/config_schemas.md) explicitly set or recommend trust_remote_code=True for vLLM and other frameworks. This setting allows the execution of arbitrary Python code bundled with models downloaded from remote repositories like HuggingFace.
[DATA_EXFILTRATION]: The presence of a generic file-copying mechanism in generated templates creates a surface for data exposure. An attacker could potentially use indirect prompt injection to trick the agent into using these flags to copy sensitive system files (e.g., SSH keys or environment secrets) to a directory where they could be easily accessed or further exfiltrated.
[PROMPT_INJECTION]: The skill uses a template-based code generation logic in scripts/generate_all.py that relies on simple string replacement (t.replace("{" + k + "}", v)) without sanitization. This is a vulnerability surface for indirect prompt injection, where a malicious user could inject code into the framework names or other parameters that would then be written into the executable generated scripts.

agent-backend-adapter