High-Risk MCP Tools

High severity 21,510 tools 3,512 servers

High-risk MCP tools trigger processes with real-world side effects. A single call can start expensive compute jobs, send mass notifications, or kick off multi-step workflows. Rate limits and argument validation are essential; approval for expensive paths is recommended.

Attacks targeting high-risk tools

Named attack patterns where tools at this severity have produced real incidents. Each links to the full case and the defensive policy.

Browse the full MCP Attack Database →

Servers with high-risk tools

Showing 50 of 3512 servers. Each server link opens its capability-level browse; each tool opens its profile with the recommended policy.

See all tools in execute.

Other risk levels

Frequently asked questions

What makes a tool high risk?

High-risk MCP tools are Execute-class: they trigger processes or run actions with side effects that reach beyond the immediate call. A build starts running, a notification fan-out begins, a workflow executes multiple downstream steps. The tool returns before the side effects finish, which makes runaway loops particularly dangerous.

What policy pattern do I need for Execute tools?

Rate-limit at two layers: per-tool call count and per-session session duration. Add argument validation so malformed inputs fail at the proxy, not inside the server. For expensive paths (CI runs, mass sends), require human approval.

How does this differ from critical-risk tools?

Critical tools are irreversible after one call. High tools are reversible or reversible-with-effort but can be expensive. The exposure is multiplicative: one runaway agent triggering 1,000 notifications each hour is more dangerous than one destructive call you can recover from.

What attacks target Execute tools?

Runaway tool loops dominate the high-risk category — self-sustaining loops burn through API quotas and third-party billing. Destructive action autonomy spills over when Execute tools chain into Destructive ones. Prompt injection via tool results compounds the problem.

What is the recommended rate limit?

Default to per-agent rate limits of 10/minute for destructive-adjacent Execute operations, 60/minute for benign ones. Session-wide call-count budgets prevent multi-hour loops. For coordinating multi-agent systems, add round-trip limits to break infinite ping-pong patterns.

Attacks targeting high-risk tools

Servers with high-risk tools

Other risk levels

Frequently asked questions

Let agents act without letting them run wild.