ZeroDrift launches three models for real-time AI compliance checks
SiliconANGLE Paul Gillin
ZeroDrift launched three AI models that check messages before they go out. It’s built for compliance teams that can’t read every bot reply by hand.
Based on reporting by SiliconANGLE, Paul Gillin — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
ZeroDrift has put a new set of compliance models into general release, betting that AI systems can police AI systems before the lawyers have to clean up the mess. The startup’s Anchor 3.0 family is meant to inspect messages generated by AI agents in real time, then flag, rewrite, block, or send them to a human reviewer if they cross a line.
That matters because autonomous agents can churn out customer messages far faster than a compliance team can review them. ZeroDrift says the models can enforce financial rules and internal company policies while still moving quickly enough to check every outgoing message before it leaves.
The company is also making a big claim about speed and accuracy. In a benchmark built around FINRA rules, ZeroDrift said its flagship model caught more than 90% of violations and matched the overall accuracy of 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. Those numbers come from a benchmark ZeroDrift published itself, even though the test data was labeled by Surge AI Inc., so they remain company claims.
Anchor 3.0 comes in three versions. Mini is a 9 billion-parameter mixture-of-experts model with 4 billion active parameters, tuned for high-volume traffic and prebuilt rule packs. ZeroDrift says it found 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 main Anchor 3.0 model uses the same parameter count but can apply more than 200 prebuilt rules covering FINRA, the Securities and Exchange Commission and other regulations, and it can identify the exact lines that break a rule before rewriting them.
Then there’s Anchor 3.0 Max, a 27 billion-parameter model aimed at long documents, attachments and a company’s own policies without extra fine-tuning. All three were post-trained from Google LLC’s Gemma E4B and Alibaba Group Holding Ltd.’s Qwen3.8-27B models. The release follows an August preview, where ZeroDrift described Anchor as a risk-reduction layer rather than a guarantee that every violation would be caught. That’s the sensible line, honestly: compliance software that pretends to be perfect usually ends up being marketing with a dashboard.
My take — AI-written commentary, not fact-checked reporting
This is the right direction, and also a very familiar one: the industry keeps inventing faster ways to make more messages, then sells software to stop those messages from becoming a problem. ZeroDrift’s pitch sounds less like hype than a necessary patch for agent sprawl. The real test is boring and brutal: whether the checks stay useful when nobody is watching the demo.
Read more about this at: SiliconANGLE