An agent with tool access is only as safe as the checks running underneath it. This directory covers the tools built to sit in that gap — watching what an agent is about to do, and stopping malicious tool use, prompt injection, and policy drift before it executes.
The tools below combine these three capabilities in different proportions — runtime interception is what makes the other two actionable rather than just informative.
Passive discovery across network, cloud, and code — surfacing shadow AI agents nobody registered with security, without requiring the agent to self-report.
Inspect not just what tools an agent calls, but why — analyzing reasoning traces alongside activity logs to catch misaligned or out-of-remit behavior early.
Check every tool call against policy at the execution boundary — approve, escalate for review, or block outright — with an audit trail for every decision.
Four tools spanning open-source runtime monitoring, policy-based tool-call enforcement, and enterprise-wide agent discovery.
Open-source and free forever: analyzes both an agent's activity logs and its reasoning traces to catch malicious tool use, prompt injection, and policy drift in real time, before the agent acts. SDKs for Python, LangChain, Anthropic, TypeScript, and a native Claude Code plugin.
Its Thoth platform checks every agent tool call against policy before it runs, enforcing contextual authorization at the execution boundary — approving, escalating, or blocking actions with a full audit trail.
Continuous discovery, visibility, and runtime enforcement across agents, tools, and prompts — detecting shadow AI, preventing data leakage through prompts, mapping agent permissions, and blocking high-risk actions at execution time.
Passively discovers and monitors every AI agent across network, cloud, and code sources — no code changes or agent self-registration required — then enforces controls for complete visibility and continuous control of shadow AI.