ZeroDrift Inc., a startup that automates compliance for artificial intelligence communications, today launched Anchor 3.0, a family of small language models designed to check messages generated by AI agents before they’re sent.
The company said the models can enforce financial regulations and internal corporate policies while operating fast enough to examine every outgoing message in real time. Anchor 3.0 is generally available through ZeroDrift’s Enforcement application programming interface.
The approach targets a growing operational problem: Autonomous agents can generate thousands of customer communications faster than human compliance teams can review them.
ZeroDrift reported that its flagship model detected more than 95% of violations in a benchmark based on Financial Industry Regulatory Authority rules. It said the model matched the overall accuracy of OpenAI Group PBC’s GPT-6 Astra and Anthropic PBC’s Claude Fable 5.1 while operating more than 34 times faster and at 1/12th the cost.
The benchmark used attorney-labeled data produced independently by data-labeling company Surge AI Inc., according to ZeroDrift. However, ZeroDrift published the benchmark itself, so the performance figures remain company claims.
“Frontier models made it easy to build capable agents,” founder and Chief Executive Kumesh Aroomoogan said in a statement. “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.”
Anchor operates within ZeroDrift’s broader compliance platform, which intercepts communications and checks them against regulatory requirements and company policies. The system can flag, rewrite, block or route problematic messages for human review while recording the decision for later audit. The company recently introduced Guard for Agents, an API-based service that inserts those checks directly into agent workflows.
The new family has three versions. Anchor 3.0 Mini is a 9 billion-parameter mixture-of-experts model with 4 billion active parameters. Intended for high-volume traffic, it runs prebuilt rule packs and flags violations. ZeroDrift said Mini caught about 5% more violations than Claude Fable 5.1 and about 20% more than GPT-6 Astra in its FINRA test, with fewer than half as many false positives as Claude.
The flagship Anchor 3.0 model has the same parameter count but can apply more than 200 prebuilt rules covering FINRA, the Securities and Exchange Commission and other regulations. It identifies the lines that violate a rule and can rewrite them before a message is released.
Anchor 3.0 Max is a 27 billion-parameter model intended for long documents, attachments and enforcement of a company’s own policies without additional fine-tuning. The models were post-trained from Google LLC’s Gemma E4B and Alibaba Group Holding Ltd.’s Qwen3.8-27B models.
The production release follows an August preview in which ZeroDrift disclosed that Anchor combined deterministic checks with an open-source model trained on regulatory material and attorney-labeled communications. At that time, the company described the software as a risk-reduction layer rather than a guarantee that every violation would be caught.
