NeuralTrust | The leading security platform for generative AI

Agentic systems add an attack surface that prompt‑only defenses miss: the agent’s own system prompt and tool/function definitions. A poisoned MCP tool description can carry an injection that never appears in the user’s message. tool_guard inspects that definition and is scoped to the mcp protocol.

Detector	Slug	Sides	Protocols	Backend
Tool Guard	`tool_guard`	input	mcp	NeuralTrust Firewall

Tool Guard — `tool_guard`

Scans the agent’s own definition — system prompt and tool/function descriptions (including parameter‑schema descriptions) — for jailbreaks and prompt injections planted there, scored by the NeuralTrust Firewall jailbreak detector (signal.type: "jailbreak").

Settings

Field	Type	Required	Notes
`jailbreak.threshold`	number	✅	Score in `[0, 1]` above which a finding is reported.
`credentials.*`	object	—	Override the global NeuralTrust Firewall credentials for this detector.

{
  "name": "Guard MCP tool definitions",
  "plugin_slug": "tool_guard",
  "settings": { "jailbreak": { "threshold": 0.6 } }
}

Scope the detector to MCP traffic and choose Monitor or Block in the policy rule (with a protocol eq mcp condition if the collector also carries other traffic). See Policies.

When to use

Whenever you load third‑party or user‑supplied MCP servers / tool descriptions — it catches injections hidden in tool metadata before the agent acts on them.
Pair it with content‑security detectors on the user’s prompt and the model’s output for defense in depth.

​Tool Guard — tool_guard

​Settings

​When to use

Tool Guard — `tool_guard`

Settings

When to use