Glossary
Threshold
The score cut-off at which a guard’s probability turns into an action, such as asking a human when P(unsafe) is at least 0.3.
Checks run on our GPUs in Switzerland. Request payloads are not stored.Security
What Threshold means
A threshold converts a continuous score into a decision. Lower thresholds catch more unsafe actions but interrupt or block more safe ones; higher thresholds do the reverse. There is no universally right value; it depends on the cost of a missed harmful action versus the cost of a false alarm in your setting.
MCP Guard’s default verdict uses two thresholds on P(unsafe): 0.8 for block and 0.3 for ask, plus 0.5 on args_grounded for ask. You can apply your own instead, and different tools can have different thresholds: stricter for payments or production writes, looser for reads.
A threshold chosen on one dataset may not behave the same on your traffic (see distribution shift). The dependable way to set one is to label a few hundred of your own agent actions, run them through the guard, and pick the cut-offs that give the precision and recall you need.
Related terms
How Threshold connects to the rest of the vocabulary.