safety-compliance — agentic threat model

Not certain from the listing — The plugin likely relies on Claude models to evaluate compliance checks. Threats include adversarial prompt injections designed to bypass approval guards or trick the model into misclassifying harmful actions as safe.

Not certain from the listing — It is unclear how policy rules or stateful circuit-breaker data are stored. Threats include tampering with local configuration files or state stores to disable active kill switches.

The plugin integrates directly into the orchestration framework (Claude Code) as hooks gating agent actions. Threats include hook bypass vulnerabilities, race conditions where actions execute before the guard resolves, and insecure tool integration.

Not certain from the listing — The plugin runs within the host environment of the parent agentic system. Threats include local privilege escalation if the plugin or its approval commands run with elevated system permissions.

This plugin acts as an active guardrail and observability mechanism. Threats include blind spots where certain agent actions bypass the hook registration, or failure to log denied actions, hindering post-incident forensics.

This is the primary layer of the plugin, enforcing authorization policies and kill switches. Threats include logical flaws in the approval gate logic, policy misconfigurations, and the risk of a single point of failure disabling all compliance checks.

The plugin ships agents and commands specifically to enforce approval gates across multi-agent systems. Threats include compromised downstream agents spoofing approval signals or cascading failures if a global kill switch is triggered maliciously.