Optimus Consulting
The Week in AI
Edition 07
UK / Service SMEs
This week, securing AI mattered more than shipping it.
The platforms spent the week adding locks, not just features. OpenAI gave ChatGPT a Lockdown Mode. Attackers turned Meta's own AI support agent against its users. For any business letting AI act on its behalf, the message is plain. Autonomy and security now have to ship together.
Top stories
5 to knowOpenAI ships Lockdown Mode as agents come under attack.
OpenAI introduced Lockdown Mode for ChatGPT, restricting internet access, file downloads and agent features to cut prompt-injection risk. Days earlier, attackers abused Meta's AI support agent to hijack Instagram accounts. Security is becoming a core setting, not an afterthought.
Optimus angle
Before you switch on any agent that can act, write down what it must never touch. Customer data, payments, the send button. Guardrails first, autonomy second.
Microsoft Build puts always-on agents on the desktop.
At Build, Microsoft pushed Scout and Autopilot agents that handle routine coordination across Teams, Outlook, OneDrive and SharePoint, plus a standalone Copilot desktop app and an image model inside PowerPoint. The pitch is an AI coworker, not a chat box.
Optimus angle
An always-on agent is only useful if you can see what it did. Turn on the activity log before the agent, so every action it takes is something you can review on Monday morning.
OpenAI turns Codex into a tool for building tools.
OpenAI rolled out role-specific Codex plugins across sales, analytics, design and finance, and Codex Sites can now build, deploy and host working apps from a prompt. Codex is moving from a coding tool to a way to spin up internal tools without a developer.
Optimus angle
The barrier to building a small internal tool just dropped. Pick one clunky spreadsheet your team fights every week and test whether a prompt-built tool can replace it before you buy software.
Anthropic heads for an IPO and warns on the brakes.
Anthropic reported annualised revenue near 47 billion dollars and confidentially filed for an IPO, even as it urged rivals to consider slowing frontier development over fears that AI could soon improve itself. Money and caution arrived in the same breath.
Optimus angle
When the people building the technology talk about brakes, take governance seriously. Decide now who in your business signs off on a new AI tool before it touches a customer.
Bots overtook humans as the web's main traffic.
Cloudflare said automated bots now make up the majority of internet traffic, driven by AI crawlers and agentic systems. The web is increasingly machines talking to machines, which changes how your site is read, scraped and represented.
Optimus angle
Assume an AI now reads your website before a human does. Check that your key pages state plainly who you are, what you do and who you serve, in language a machine can quote back.
From the Insurance Desk
Sector signal for brokers, motor claims and credit hire companies.
Item 01
QBE warns nobody is checking their suppliers' AI.
QBE flagged that most firms do not audit the third-party AI systems they rely on, opening a supply-chain cyber and coverage gap. For brokers and claims firms leaning on vendor tools, an unaudited supplier is now an unaudited risk.
Optimus angle
List every outside tool that touches your claims or customer data and uses AI. Ask each supplier one question. How do you secure it. No clear answer is itself an answer.
Item 02
Hiscox starts paying to appear inside ChatGPT.
Hiscox UK joined an early OpenAI ads pilot through Dentsu, putting marketing budget into being seen inside the chatbot. It is the sixth insurance move into ChatGPT in four months, and the first about paid advertising rather than quote apps. The buying journey has moved.
Optimus angle
You will not outbid Hiscox on a new ad surface, so do not try. Earn the slot instead. Be the clearest, best-evidenced answer when a customer asks the engine for a recommendation. That is where smaller brokers and claims firms still win.
Editor's column
The Optimus Take
The week's headlines were about locks, not launches. OpenAI added a Lockdown Mode. Meta watched its own support agent get turned against its users. QBE pointed out that hardly anyone checks the AI sitting inside their suppliers. The pattern is clear. As AI starts to act, the question shifts from what can it do to what could go wrong.
For UK service SMEs this is good news, oddly. You do not need a frontier model to be safe. You need to know which tools can act on their own, what they can reach, and who approved them. That is an operations job, not a technology one. It is also something you can start this week without spending a penny.
- StrategyDecide which AI tools are allowed to act on their own inside your business, and which must always ask a human before they do.
- OperationsTurn off every agent feature you are not actively using. Autonomy off by default, on by request, logged when in use.
- ServiceName the person who approves any new AI tool before it touches customer data or customer communications. One owner, one sign-off.
“Give AI the keys to your work, and you own the locks.”