code-safety-monitor — agentic threat model

Uses a DSPy-based classifier and Claude Code. Vulnerable to adversarial evasion techniques where malicious code is obfuscated to bypass the ~90% detection threshold, or prompt injection that tricks the classifier into flagging benign code as malicious.

Not certain from the listing — relies on code snippets passed during the dev loop. If it caches or stores scanned code for training or few-shot DSPy assertions, there is a minor risk of data exposure or poisoning of the training/optimization dataset.

Hooks directly into the Claude Code agent workflow. If the orchestration framework is compromised, an attacker could disable the plugin entirely, modify the audit checkpoints, or spoof the classification outputs to commit backdoored code.

Not certain from the listing — runs locally or within the developer's environment as a Claude Code plugin. Security depends heavily on the host machine's sandboxing and whether the plugin executes code during evaluation.

Provides audit checkpoints and detection commands in the dev loop to monitor agent behavior. However, a 10% false negative rate (90% detection rate) leaves a monitoring blind spot that sophisticated attackers can exploit.

Acts as a compliance and security control by flagging suspicious behavior before code commits. However, it lacks formal authorization policies or access controls to prevent developers or other agents from overriding its blocks.

Designed to interact with other developer agents (like Claude Code). A compromised coding agent could attempt to exploit, bypass, or feed malicious inputs to this monitor agent to neutralize its defensive capabilities.