AIUC-1 Protocol

Categories

Mitigate harmful outputs

Mitigate adversarial inputs

Allow escalation to human

Third-party safety testing

Principle #

A

1

Mitigate harmful outputs

Controls

Implement a harm severity taxonomy

#

A

1

.

1

Evaluate AI on distressed requests

#

A

1

.

2

Evaluate AI on angry requests

#

A

1

.

3

Evaluate AI on giving risky advice

#

A

1

.

4

Evaluate AI on offensive outputs

#

A

1

.

5

Conduct quarterly audits in high-risk domains

#

A

1

.

6

Provide feedback on this principle