ZeroDrift on Wednesday, September 23, 2026, announced the general availability of Anchor 3.0, a family of small language models designed to check AI-generated communications against regulatory and company rules before they are sent.
The New York-based company said the models are built for real-time enforcement across AI agents, applications and enterprise workflows. Developers can access Anchor 3.0 through ZeroDrift’s Enforcement API.
ZeroDrift’s system sits between an AI model and the person or system receiving its output. It can pass a message, identify a violation, rewrite AI-generated content, block a message or escalate it for review.
The launch comes as financial institutions increase their use of AI. Research from the Cambridge Centre for Alternative Finance found that 81% of surveyed financial services firms were adopting AI at some level, while 52% reported active adoption of agentic AI. Among traditional financial institutions, active agentic AI adoption stood at 45%.
FINRA’s 2026 regulatory oversight report says its rules and securities laws continue to apply when firms use generative AI. The regulator specifically identifies supervision, communications, recordkeeping and fair-dealing obligations as areas that can be affected by the technology.
Anchor 3.0 is available in three versions. Anchor 3.0 Mini has 9 billion parameters, with 4 billion active parameters, and is designed for high-volume traffic and violation detection.
The flagship Anchor 3.0 also has 9 billion parameters and 4 billion active parameters. It was post-trained from Gemma E4B and is designed to run more than 200 pre-built rules covering FINRA, the Securities and Exchange Commission and other regulations.
The flagship model identifies the specific lines that violate a rule and can rewrite AI-generated content before delivery. ZeroDrift said customer-specific policies can be trained into the model through a LoRA adapter.
Anchor 3.0 Max has 27 billion parameters and was post-trained from Qwen3.8-27B. It is designed for longer documents, attachments and enforcing company policies without additional fine-tuning.
ZeroDrift said Anchor combines a purpose-built enforcement model with a deterministic rules engine. Both evaluate messages and are reconciled into a single decision.
The system also rechecks rewritten content before delivery. Its decisions can identify the regulation or company policy behind an enforcement action.
For its launch, ZeroDrift published a FINRA benchmark based on 150 human-written communications. The dataset contained 98 communications with deliberate FINRA violations and 52 clean communications.
ZeroDrift said the communications were independently produced by Surge AI and labelled by attorneys. Three independent reviewers assessed each item, with 95% agreement on the verdict and 93% agreement on the intended rule.
In the benchmark, ZeroDrift reported that Anchor 3.0 caught 95.5% of violations, the highest recall among the frontier models it tested. The company also reported that Anchor outperformed GPT-5.6 Sol on recall, precision and F1, while running up to 34 times faster and up to 12 times cheaper.
ZeroDrift reported processing times of about 1.5 seconds per message through its API and less than 100 milliseconds when self-hosted in a customer’s virtual private cloud environment.
The company’s benchmark also reported a cost of $10 per 1,000 messages for Anchor 3.0, compared with $48 for GPT-5.6 Sol in the same comparison.
SiliconANGLE reported that ZeroDrift’s flagship model detected more than 95% of violations in the FINRA test and that the benchmark was published by ZeroDrift. The publication also reported that Anchor 3.0 matched the overall accuracy of GPT-6 Astra and Claude Fable 5.1 in the company’s test while operating substantially faster and at lower cost.
Anchor is part of a broader ZeroDrift platform that includes Guard for Agents and Command. Guard for Agents, introduced on September 3, places the enforcement check directly into an agent workflow through a production Validation API.
Guard also provides an OpenAPI specification and MCP connectors for Notion, Linear and Google Drive. ZeroDrift said those connectors can turn existing policy documents into enforceable rules.
Command serves as the control centre for managing policies, testing enforcement and reviewing activity. ZeroDrift records enforcement decisions, including the original message, the rule triggered, the action taken and the output that ultimately passed through the system.
ZeroDrift launched in 2026 and is headquartered at 4 World Trade Center in New York. The company was founded by Kumesh Aroomoogan, who previously co-founded Accern, a financial-services AI company that was acquired in 2025.
The company emerged from stealth in February with a $2 million pre-seed round led by a16z speedrun. At launch, ZeroDrift said it was initially targeting registered investment advisers, asset managers, broker-dealers and wealth platforms.
In June, ZeroDrift announced a $10 million seed round involving a16z speedrun, Reign Ventures, PitchDrive Ventures, U&I Ventures, Active Capital, Geek Ventures, Converge Ventures, Atlas SGR, Founders Future and other investors.
That brought the company’s publicly announced fundraising to $12 million. TechCrunch reported that the June seed round was closed in three weeks and was oversubscribed by three times, according to Aroomoogan.
ZeroDrift said the platform has gained traction with tier-one banks, asset managers and insurance companies. The company has also said that its broader plans include expanding enforcement across insurance, healthcare and other regulated sectors.
The company says its platform supports regulatory requirements including SEC and FINRA rules, as well as frameworks such as MiFID II, GDPR and HIPAA. Its published security information lists SOC 2 Type II, ISO 27001, GDPR and HIPAA, along with single sign-on, virtual private cloud deployment, per-tenant isolation and customer-managed encryption keys.
ZeroDrift is charging $0.01 per enforcement for the flagship Anchor 3.0 model, according to its published benchmark materials. Anchor 3.0 Mini is listed at $0.002 per enforcement, while Anchor 3.0 Max is listed at $0.05 per enforcement.
New accounts are currently being offered up to $1,000 in complimentary credits through September 30, 2026, according to ZeroDrift’s platform pages. The company says the credits expire on December 31, 2026.
ZeroDrift’s latest release places a specialised compliance model directly in the path of AI-generated communication, with the company positioning Anchor 3.0 as the enforcement layer between enterprise AI systems and the people or systems receiving their output.
Further information is available in ZeroDrift’s Anchor 3.0 benchmark and methodology, the company’s Guard for Agents documentation, FINRA’s 2026 guidance on generative AI, and the Cambridge 2026 Global AI in Financial Services Report.
Discover more from Aree Blog
Subscribe now to keep reading and get access to the full archive.
