AI red teaming: what a test shows, and what it does not
Red teaming asks what a system does under pressure. We fix the version, the configuration and the test set before the run, so the result can be checked rather than believed.
A red team report is useful only if it can be reproduced. We publish the protocol next to the findings, including the runs that found nothing. A negative result is a result, and it is recorded in the same way as a breach.
Most failures we see are not exotic. An assistant follows an instruction hidden in a document; a tool call leaves the boundary it was given; a guardrail falls to a simple paraphrase. Each of these becomes a finding with a trace, not an opinion.
What red teaming does not do: it does not certify safety and cannot prove absence of risk. It shows what happened under fixed conditions. That limit is written into every report we issue.