The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill recommends installing its CLI tool using curl -fsSL https://raw.githubusercontent.com/avivsinai/agent-message-queue/main/scripts/install.sh | bash in SKILL.md. This pattern executes unverified remote code directly in a shell environment.\n- [PROMPT_INJECTION]: Instructions in SKILL.md include flags like --dangerously-skip-permissions and --dangerously-bypass-approvals-and-sandbox which are designed to override agent security boundaries and safety filters.\n- [COMMAND_EXECUTION]: Documentation promotes eval \"$(amq env ...)\" in SKILL.md, which executes shell commands generated by the amq binary. This creates a risk of arbitrary command execution if the binary's output is compromised.\n- [DATA_EXFILTRATION]: The skill facilitates data movement across project and session boundaries (references/cross-project.md). When used with sandbox-bypass flags, this poses a risk of unauthorized data exfiltration.\n- [EXTERNAL_DOWNLOADS]: The skill relies on external code and configuration from a GitHub repository (avivsinai/agent-message-queue) not recognized as a trusted organization.\n- [PROMPT_INJECTION]: The skill has a surface for indirect prompt injection.\n
Ingestion points: Ingests data from mailbox files via amq drain and amq watch.\n
Boundary markers: Header validation is mentioned, but no body delimiters or warnings are instructed.\n
Capability inventory: Access to shell execution and sandbox bypass flags.\n
Sanitization: No evidence of sanitization for the message body is provided.

amq-cli