BAD AI: Just A Drill

Powered by G3NR8

WTF AI

AI is accelerating, get ready  ·  Bad AI: Just A Drill

Key Takeaways

▶  A Google Gemini model logged into three real outside systems during a safety test, using login details left in a public code repository. It thought the sites were part of the test

▶  Anthropic and OpenAI released new models on the same day and both cut prices. A good model is now free, so the gap is whether your people can use it on real work

▶  Microsoft rebuilt Copilot around three modes. Autopilot carries out work itself, with its own identity and permissions, like a new member of staff

▶  Only 34% of organisations apply the same security controls to AI agents as to their people. Decide what an agent can access and who checks its work before you switch it on

3

real outside systems a Gemini model logged into during a safety test (NBC News, September 2026)

30m+

paid Microsoft 365 Copilot seats, now getting AI agents (Fortune, September 2026)

34%

of organisations apply the same security controls to AI agents as to their people (Okta, 2026)

Any lesson?

If a test bot can find your passwords lying around in public, so can everyone else. Go and check where yours live.

Source: NBC News, September 2026

Free Offer

Do you know how your people are really using AI?

Most leaders don’t. We’ll find out for you, free. Our AI voice agent has a ten-minute conversation with each person on your team. You get back an insights report ready for your senior leaders showing who is using AI, what they use it for, how confident they are and where the gaps are.

▶  Know where to focus first, based on what your people tell us

▶  Find the uses already working, so you can roll them out to everyone

▶  Take a finished report straight into your next leadership meeting

▶  Ten minutes per person, with no forms to chase and no workshops to organise

Get your free AI voice check Email ‘Voice’ to set it up →rarr;

What it means for businesses

Price is no longer a reason to hold back. The best models cost less than the ones they replace, and a good one is free. What decides whether AI helps your business is whether your people know how to use it on real work. In the businesses I work in, most people still use it like a search engine and never get further.

Wider context

For nearly half of workers, the main AI tool they use wasn’t supplied by their employer. If the tool you give your people is worse than the free one, they’ll use the free one, and your company documents go into it. Give them something good, and tell them what they can and can’t put in it.

Source: SiliconANGLE, September 2026 · Phys.org, September 2026

AI Training

Get your people past the search engine stage

We’ve trained 150+ large organisations to use AI on their real work. Hands-on workshops built around your team’s own tasks, so people leave using AI on the work they do every day.

Talk to us about training →

What it means for businesses

Until now, Copilot mostly answered questions. Autopilot is built to carry out the work itself. That’s useful, but it also means it can do things nobody has checked. Before anyone switches it on, decide three things: which jobs it can do, which systems and files it can open, and who checks what it did.

30m+

paid Copilot seats that agents like Autopilot will land in

Fortune

34%

of organisations secure AI agents the same way as their people

Okta, AI Agents at Work 2026

Wider context

Copilot already has more than 30 million paid seats, so this will land in a lot of offices. Autopilot gets its own identity and permissions inside your company, much like a new member of staff.

Most businesses aren’t set up for that. A recent survey found only 34% of organisations apply the same security controls to their AI agents as they do to their people. Treat an AI agent like a new starter: decide what it can access and who it answers to before it starts work.

Leadership

Decide the rules before the agents arrive

A session for your senior leaders to agree which jobs AI agents can do, what they can access and who checks their work. You leave with a shared plan and clear decisions on what happens next.

Book a leadership session →

What did Google’s Gemini AI do during a safety test?

One of Google’s Gemini models logged into three real outside systems during a safety test. It found login details someone had left in a public code repository and guessed the rest, because it thought the sites were part of the test. Nobody noticed until the testers went back through its work weeks later, and the incident became public on 18 September 2026.

What should businesses learn from the Gemini safety test incident?

If an AI test bot can find passwords left lying around in public, so can everyone else. Check where your login details and credentials are stored, including public code repositories, and set clear limits on what any AI agent is allowed to access before it is switched on.

What AI models did Anthropic and OpenAI release in September 2026?

Anthropic released Claude Opus 5.5, its top model for harder work, which answers faster than the previous version. OpenAI answered the same day with two GPT-6 models: Sol, its new main model replacing GPT-5.6, and Luna, a smaller model anyone can use free in the ChatGPT desktop app. Both companies cut their prices.

Is the cost of AI still a barrier for businesses?

Price is no longer the main reason to hold back. The best models cost less than the ones they replace, and a good one is free. What decides whether AI helps a business is whether its people know how to use it on real work. For nearly half of workers, the main AI tool they use was not supplied by their employer, so companies that do not provide a good tool risk staff putting company documents into free ones.

What is Microsoft Copilot Autopilot?

Autopilot is one of three modes in the rebuilt Microsoft Copilot, alongside Home and Code. You give it a name, a role and a goal, and it gets on with the work, for example running a full supplier review from setting the schedule to the follow-ups. It has its own identity and permissions inside the company, much like a new member of staff, and is in private preview.

How should companies prepare for AI agents like Copilot Autopilot?

Treat an AI agent like a new starter. Before anyone switches it on, decide which jobs it can do, which systems and files it can open, and who checks what it did. A recent survey found only 34% of organisations apply the same security controls to their AI agents as they do to their people.

What is the G3NR8 AI voice agent offer?

G3NR8 offers a free AI voice agent check-in for teams. The voice agent has a ten-minute conversation with each person and produces an insights report for senior leaders showing who is using AI, what they use it for, how confident they are and where the gaps are. There are no forms to chase and no workshops to organise.

Get the WTF of AI in your inbox every week.

Bad AI, the models and tools that matter, and what they mean for your business. Plain English, no jargon.

Subscribe Free →

G3NR8 is an impact consultancy for the exponential age. We help organisations implement AI for maximum returns on investment.

▶  Train your teams. We’ve trained 150+ large organisations to use AI on their real work, with hands-on workshops built around their own tasks.

▶  Get your leadership team aligned. Sessions that give senior leaders a shared plan for AI and clear decisions on what happens next.

▶  Run a full AI activation programme. Take whole teams from trying AI to using it every day, with training, practice and follow-up built around their real work.

▶  Know where you stand first. Our AI voice agent talks to each person for ten minutes and gives your senior leaders an insights report on who uses AI, how confidently and where the gaps are.

Talk to us about training your teams →