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.

Strength is tested by load, not by promise Strength is tested by load, not by promise
Strength is tested by load, not by promise

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.

Materials on this topic

Other topics

AI standards: where the rules actually applyAI supply chain: what enters before the model runsMulti-agent systems: behaviour between the agentsIncident chronicle: documented consequences, not forecastsBehavioural audit: how agent behaviour drifts over timePrompt injection: where text becomes an instructionRussian practice: who is held responsible for AIMethodology: how the Institute tests AI systemsData and benchmarks: what our conclusions rest onSociology of AI: relations between the agentsGovernment customers: how to commission a testGovernance and funding of the InstituteCases: scenarios of work and what they show

Page updated 02 November 2026 · site build 2026.11.13 · ANO «ISAIS»