ZeroDrift Releases Anchor 3.0 Family to Automate AI Compliance Screening

ZeroDrift Inc. has introduced Anchor 3.0, a suite of compact language models built to monitor communications produced by AI agents before delivery. The models enforce both regulatory requirements and internal business rules while maintaining the speed necessary to inspect every message as it flows through systems in real time. The offering is now available via ZeroDrift's Enforcement application programming interface.

The product addresses a mounting challenge for regulated industries: autonomous agents can produce thousands of customer messages faster than human compliance reviewers can keep pace. ZeroDrift's models aim to close that gap by automating the review process at scale.

In testing based on Financial Industry Regulatory Authority standards, ZeroDrift's primary model detected more than 90% of violations. The company claims the model achieves accuracy matching OpenAI Group PBC's GPT-6 Astra and Anthropic PBC's Claude Fable 5.1, while running more than 100 times faster and costing less than one 500th as much. The benchmark relied on attorney-labeled data from Surge AI Inc., though ZeroDrift published the results itself, meaning the performance claims remain unverified by independent sources.

Frontier models made it easy to build capable agents. The hard part is running them inside a regulated business, where every message has to follow the rules and the check has to happen every time, before anything goes out.

Kumesh Aroomoogan, founder and Chief Executive of ZeroDrift

ZeroDrift's compliance system intercepts outgoing communications and evaluates them against regulatory standards and company-specific policies. The platform can flag problematic content, suggest rewrites, prevent transmission, or escalate messages to human reviewers, while maintaining records of each decision for audit purposes. The company recently rolled out Guard for Agents, an API service that embeds these checks directly into agent operations.

Three Models for Different Use Cases

Anchor 3.0 Mini operates as a 9 billion-parameter mixture-of-experts architecture with 4 billion active parameters. Designed for high-volume message streams, it applies prebuilt rule sets and identifies violations. In ZeroDrift's FINRA benchmark, Mini caught roughly 5% more violations than Claude Fable 5.1 and approximately 20% more than GPT-6 Astra, while generating fewer than half the false positives of Claude.

The standard Anchor 3.0 model shares the same parameter structure but supports application of more than 200 prebuilt rules spanning FINRA, Securities and Exchange Commission and additional regulatory frameworks. Beyond detection, it can pinpoint the specific text causing violations and rewrite problematic passages before release.

Anchor 3.0 Max, a 27 billion-parameter variant, handles longer documents and attachments while enforcing organization-specific policies without requiring additional fine-tuning. All three models were built through post-training on Google LLC's Gemma E4B and Alibaba Group Holding Ltd.'s Qwen3.8-27B foundations.

The full release builds on an August preview where ZeroDrift demonstrated Anchor combining rule-based checks with an open-source model trained on regulatory texts and attorney-reviewed communications. At that earlier stage, the company positioned the technology as a risk-reduction mechanism rather than a complete guarantee against missed violations.