Red-Teaming Frontier Models in 2025
Manual red-teaming does not scale. Continuous, automated adversarial pipelines are the new baseline.
Manual red-teaming gave us the first real picture of how frontier models fail. It does not scale to weekly model releases.
The teams shipping safely now run automated adversarial pipelines on every checkpoint. Humans review the outputs, not the inputs.
What changed in the last year
LLM-generated jailbreaks have closed the gap with skilled human attackers. They explore the attack surface faster and remember every prior attempt.
Coupled with model-graded evaluation, this lets a small team monitor coverage that previously required dozens of red-teamers.
Operational guidance
Treat red-teaming as a CI step, not a release-gate ritual. Every PR that touches a prompt or a tool deserves an automated adversarial pass.
If your red-team only runs before a launch, your launches are the only safe moment.
