ZeroDrift launched Command on August 11, a compliance control plane that checks human- and AI-generated communications before delivery using deterministic rules and the company's Anchor 3.0 Preview small language model. The company says the system can pass, warn, repair or block messages, but it has not published detailed detection or false-positive results for independent review.
ZeroDrift launched Command on August 11 as a control plane for checking human- and AI-generated communications before they are delivered. The system combines deterministic policy rules with Anchor 3.0 Preview, a purpose-built small language model, to identify potential regulatory and internal-policy violations.
How the control plane works
SiliconANGLE reports that Command is organized around four components. Policy Studio converts regulations and company policies into enforceable rules. Anchor analyzes content for potential violations. Guard applies those policies where communications enter or leave an organization, while Command records decisions and the evidence behind them.
The product can pass, warn, repair or block a message. ZeroDrift says AI-generated text may be rewritten automatically, while a human-authored message can instead be blocked, routed to a compliance officer or returned with suggested changes. The company also offers an API and email integration through an SMTP relay for Microsoft and Google systems.
ZeroDrift's current product page describes Anchor 3.0 Preview as a compliance SLM running alongside a deterministic rules engine. It says rewrites are checked again before delivery and each decision cites the rule that triggered it. The stated coverage includes financial services, insurance, healthcare and custom company policies, with cloud or private-cloud deployment options.
Performance evidence is still limited
Chief executive Kumesh Aroomoogan told SiliconANGLE that Anchor uses an open-source model post-trained on regulatory data and communications labeled by attorneys. He said ZeroDrift is evaluating it against 100,000 human-annotated emails and that the preview detects at least twice as many violations as general-purpose frontier models. The company has not released the underlying detection rates, false-positive rates, comparison protocol or dataset details, so the claimed advantage cannot yet be independently evaluated.
Aroomoogan also said the specialized model reduced detection-and-rewrite latency from more than ten seconds in earlier Claude-based approaches to under one second, sometimes milliseconds. Those are vendor-reported figures rather than published benchmarks.
For AI teams in regulated environments, the architecture illustrates a practical pattern: combine deterministic policy checks with a specialized model, retain an audit trail and revalidate rewritten output. ZeroDrift itself cautions that Command reduces risk rather than guaranteeing that every violation will be caught. Production buyers should therefore evaluate recall, false positives, policy-update latency, escalation behavior and audit evidence against their own regulated workflows.
Key Points #
- 1Command combines deterministic policy checks with the Anchor 3.0 Preview small language model before messages are delivered.
- 2The platform can pass, warn, repair or block content and records the rule and disposition for each decision.
- 3ZeroDrift has not released detailed detection, false-positive or independent benchmark results for its performance claims.
Scoring Rationale #
A current product launch with a concrete hybrid rules-and-model architecture relevant to regulated AI deployment, balanced by the absence of independently reviewable accuracy and latency benchmarks.
Sources #
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.