Definition
Plain language
A system that stress-tests AI agents for safety by checking whether harm actually happened in the world, not by trusting what the agent says it did.
As stated in the literature
An agent safety-testing framework that mines the literature into a risk taxonomy, composes thousands of reproducible attack scenarios, and judges each by inspecting observable environment state through a deterministic verifier rather than the agent's self-report.
Why it matters: It catches cases where an AI says the right thing but does the wrong thing, giving a truer picture of real-world safety.
For example, instead of accepting an agent's claim that it 'refused to send the money,' Vera checks the account balance to see if funds actually moved.