Thursday 19th November 2026
In October, Google announced AI agents that get their own work account and their own email address. Not assistants that wait to be asked, but coworkers that pick up a task and get on with it.
The UK is not far behind. Octopus Energy's assistant, Arlo, answered around 8,000 routine customer emails a week in its trial, and customers rated its replies higher than comparable human ones. A&O Shearman has built agents that work through antitrust filings and loan reviews step by step. Lloyds is planning agents for fraud investigation and complaints. Agents are moving from the demo to the day job, and quickly.
Agents do not take no for an answer
That drive comes with a trait worth understanding. In September, Australia revealed that an OpenAI agent researching medicines spending had been refused by the Medicare statistics portal, and found another way in. It "didn't accept no for an answer," said Prime Minister Anthony Albanese.
It was a test agent with its safeguards lowered, but it is not an outlier. When the UK's AI Security Institute tested five leading models on cyber tasks this summer, every one of them tried to cheat at some point, stepping outside the rules to reach its goal.
None of this is malice. Agents are built to finish the job. Give one a goal and it will use every route its access allows, without stopping to ask whether it should be looking.
What it finds is down to you
In part one we argued that data and identity are now one problem, because an agent is both a user and a piece of software. This is what that looks like in practice. The question is not whether your agents will go looking. It is what they find when they do.
That depends on your data, and in most estates the tools deciding what is sensitive work from an idea that is 75 years old. They look for patterns. Sixteen digits in a row is probably a card number. Two letters, six numbers and a letter is probably a National Insurance number. The technique, the regular expression, was set out by the mathematician Stephen Kleene in 1951.
It still works for data with a fixed shape. But a spreadsheet called Project Falcon holding the board's acquisition shortlist has no pattern to match, so it goes unnoticed. That was fine when a person had to go looking for the file. Agents go looking for you.
What we expect next
Three things we think will define the next 12 to 18 months:
Three layers of trust
The organisations that move fastest will build trust in three layers, each resting on the one below.
Get the foundation right and it stops being the brake on AI and becomes the accelerator. That is why we partnered with Cyera: it builds all three layers on one picture of your data.
We have seen this from the inside, too. Bytes now runs Cyera across its own estate, for the same reasons we recommend it to customers.
Where to start
Ask one question: what would an agent find in your estate if it would not take no for an answer?
Our Data and AI Risk Assessment can help you understand where sensitive data sits, who can access it and where unnecessary exposure may exist.
Reach out to book your assessment
To explore these challenges further, join Bytes and Cyera on 9 December for The Security Debrief: What Comes Next for Data & AI, an exclusive executive roundtable at The Ivy in the Park, Canary Wharf. We’ll explore how organisations can strengthen the data, access and security foundations needed to scale AI with confidence.
Want to keep informed? Sign up to our Newsletter