Building the Guardrails and Adversarial Testing Behind an AI Risk Approval Platform

The challenge

AetherLab's premise is that AI adoption has outrun AI approval. Nearly every AI system that touches money, customers, or regulated content eventually has to get past someone whose job is to say no, a merchant-risk team, an underwriter, a procurement reviewer, an auditor, and most of those teams have no standard way to actually test the AI system in front of them. AetherLab exists to close that gap: red-team AI systems the way an adversary would, contain what breaks with guardrails running in production, and turn both into evidence a risk team can act on.

That is a different kind of engineering problem than most AI products face. AetherLab wasn't building a chatbot feature. It was building infrastructure that other companies' risk, compliance, and procurement teams needed to trust enough to approve real deployments on, against processors, insurers, and enterprises who would stress-test it exactly as hard as it stress-tests their own AI.

What we built

Guardrails. We built PromptGuard and MediaGuard, the production guardrail layer that enforces bespoke policy on live traffic across both text and image. Every check runs against custom policy rules defined per client, at flat pricing regardless of how many rules a client runs, and every verdict returns a threat score with a written rationale rather than a bare pass or fail, so a client's logs answer "why was this blocked" before anyone has to ask.

AdversarialScan. We contributed to AdversarialScan, AetherLab's red-teaming engine, extending its adversarial testing across text, multi-turn conversations, and images rather than single-prompt text alone. Findings come back severity-scored against the specific failures that would be a real business problem for that client, not a generic jailbreak checklist, and the underlying engine models the system under test to generate adaptive attack campaigns rather than running the same static attack list against everyone.

Enterprise readiness. Verdict infrastructure that sits in the critical path of customer traffic has to survive a bad day, not just a demo. We helped build AetherLab's jury-based verdict architecture, where multiple models adjudicate every check so no single model failure or provider outage decides an outcome on its own, along with the layered degradation that keeps checks running on remaining layers when one dependency fails, rather than failing closed all at once.

Where it landed

AetherLab now processes more than 150,000 AI checks a day and screens over 17 billion tokens a month across text and image, with customers running 225+ custom policy rules each. The platform has served production traffic every hour of every day for the last 90 days running, and in head-to-head evaluations has been chosen over Amazon Bedrock Guardrails and Hive AI, including by a top-10 high-risk payment processor.

Because the guardrail layer and AdversarialScan were built to share the same underlying findings, a vulnerability AdversarialScan surfaces can be turned directly into a guardrail policy that contains it in production, closing the loop between what breaks, what it costs, what protects against it, and what the Evidence Pack proves to the risk team that has to approve it.