The Agent Skills Directory

[Credential Security Guardrails]: The skill explicitly instructs the agent to avoid echoing sensitive keys in its responses or terminal output and mandates that secrets should only be handled via .env files or platform-native secret management. This is a positive security feature designed to prevent accidental data exposure.
[Resource Isolation]: It provides guidance on using sandboxed environments for file operations and shell commands, emphasizing that these environments are isolated per thread to prevent cross-session interference.
[Secure CLI Practices]: The skill documents the use of the managed-deepagents CLI, which is installed through standard package registries (npm and PyPI/uv). It includes safety warnings for destructive commands, such as requiring user confirmation before execution.
[Data Trust Model]: There is a clear warning regarding the use of shared memory, advising that its content should be treated as untrusted input and should not be used to grant authority or bypass security approvals.
[Indirect Prompt Injection Surface]: Like most development-focused agents, this skill has an attack surface where it processes external data (such as code files, tool outputs, or shared memory). However, the instructions include explicit warnings to treat this data as untrusted, which serves as a mitigation factor.

managed-deep-agents