# Dangerous Capability Evaluation

> A test to see if an AI model could do great harm.

**Field:** Artificial Intelligence  
**Category:** Safety & Security  
**Also known as:** Dangerous Capability Eval, Frontier Risk Evaluation, Capability Evaluation (Safety)

## Think of it like

It is like a crash test for a new car. You find the danger before anyone drives it.

## Definition

A dangerous capability evaluation is a test to see if an AI model can do great harm. Teams test if it could help make germ or chemical weapons. They test if it could break into computers, or act on its own without control. Experts try hard to get the model to do these things, most often before release. The results are compared with set danger lines. If the model crosses one, stronger safeguards must be in place before it ships.

## Example

Before release, a lab asks experts to try to plan a cyber attack with a new model. It does better than the last model. So the lab adds more safeguards first.

## Why it matters

A lab cannot guard against a power it never tested for.

## Common mistake

Testing only with simple prompts. A real attacker will try much harder, so the tests must too.

## Three principles

- **Test the worst.** Weapons, hacking and escape.
- **Before release.** Find it before users do.
- **Danger lines.** Crossing one means more safeguards.

## Related terms

- [AI Red Teaming](https://learn.benibix.com/term/ai/ai-red-teaming)
- [Frontier Safety Policy](https://learn.benibix.com/term/ai/frontier-safety-policy)
- [AI Benchmark](https://learn.benibix.com/term/ai/ai-benchmark)
- [Frontier Model](https://learn.benibix.com/term/ai/frontier-model)
- [AI Safety](https://learn.benibix.com/term/ai/ai-safety)
- [Scheming (AI)](https://learn.benibix.com/term/ai/scheming)

## Told apart from

- [AI Red Teaming vs Dangerous Capability Evaluation](https://learn.benibix.com/compare/ai-red-teaming-vs-dangerous-capability-evaluation): People think red teaming and a danger test are the same check, and they are not. A red team tries to make a model break its rules, with tricks like jailbreaks and hidden orders. A danger test asks what the model can do at its strongest, such as help with a weapon or a break-in. One looks for holes in the rules. The other measures the power behind them.

![A new model goes through highlighted danger tests for weapons, hacking and escape. The result is compared with a line set in advance. If it is not past the line, the usual safeguards apply and it is released as planned. If it is past the line, stronger safeguards must be in place before release.](https://learn.benibix.com/diagrams/ai4-safety-gates.svg)

---

Source: https://learn.benibix.com/term/ai/dangerous-capability-evaluation
Author: Behnia Zarei (a developer in Vienna, Austria, who builds and teaches data, AI and business intelligence at BENIBIX)
Publisher: BENIBIX Learn (https://learn.benibix.com/)
Last updated: 2026-11-18
Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Quote freely with attribution.
