National Australia Bank is running security checks on autonomous AI systems that can make decisions without waiting for human approval. The stress-test marks a turning point for the financial sector, where banks are racing to deploy agents capable of acting independently, but nobody quite knows what happens when something goes wrong.

NAB's push into agentic AI has been relentless. In April 2026, the bank launched a dedicated AI Science team under George Mathews with a single mission: figure out how to build and evaluate autonomous systems that work inside one of the world's most heavily regulated industries. The team isn't doing abstract research. They're constructing the actual frameworks, evaluation methods, and architectural patterns that decide whether an AI agent gets released or gets shelved.

The bank already has OpenAI-based agents handling document processing, and the numbers speak for themselves. NAB processes around 15,000 trust deeds every year. Humans used to spend roughly 45 minutes reviewing each one. The AI agents do it in about 1 minute. That's a 97% cut in review time.

Across the organization, NAB has standardized AI tooling for roughly 6,000 developers and partnered with platforms like Harness to bake security and compliance checks into the development pipeline earlier. By March 2026, agentic AI applications for customer workflows had hit a 90% adoption rate across the bank's divisions.

Why autonomous AI breaks the traditional safety model

The security challenge here is fundamentally different from what banks have dealt with before. Traditional AI that hallucinates gets caught. A human sees the error before anything happens. Agentic AI doesn't work that way. It processes the hallucination and acts on it. It chains multiple steps together without pausing for approval. By the time someone notices something is wrong, the system might have already executed three moves nobody asked for.

A system running through 15,000 documents at one minute each can spread errors faster than any human team could ever catch them. For a bank NAB's size, one mistake in an autonomous workflow could touch thousands of transactions before anyone realizes there's a problem. That's why the stress-testing matters. It's not theoretical. It's the difference between controlled deployment and a cascade failure.

This article is informational and does not constitute financial or investment advice.