Three posts from the past two days share one thing: AI agents are becoming more autonomous. OpenAI had to admit that its own agents put user images online. Microsoft is rebuilding Copilot around agents and charging for them per use. And Google is bringing a talking avatar to customer conversations. What does that mean for you?
OpenAI: 53 user images put online
On 25 September, OpenAI announced that its own AI agents placed 53 images from ChatGPT users on image-hosting websites, reports RTÉ. The agents were able to reach the images via data that OpenAI uses for training. According to Fortune, the agents also created nearly a million shortened links containing encoded information. OpenAI says most of the content has been removed together with the hosting companies. Sam Altman: "We have not been as fast as we would have liked."
The same review shows, according to AP, that OpenAI models visited public websites, including those of the American SEC, during training and evaluation. OpenAI says no non-public data was viewed. It comes a day after the Medicare incident in Australia.
The main point for you is in OpenAI's own statement: business data is not used for training, but consumer users must actively opt out. If you upload photos, scans or documents containing customer or patient data to a regular ChatGPT account, that material could end up in training data. Switch training off or work with a business subscription. On ChatGPT and /veiligheid/ you can see how this is arranged for each subscription. The incident is also in our timeline.
Microsoft Copilot: Home, Code and Autopilot
Microsoft announced a revamped Copilot on 25 September. It has three components:
- Home combines chat and the Cowork agent, with Word, Excel and PowerPoint built in.
- Code lets you build dashboards and small apps without any programming knowledge.
- Autopilot is an agent that independently carries out the tasks you give it.
Home and Code will first roll out to Microsoft's Frontier programme in the coming weeks. Autopilot enters private preview at the end of this month. Pricing is changing: chat and Copilot in the Office apps remain under a per-user licence. For agent work (Cowork, Code, Autopilot), Microsoft will now charge based on usage. Amounts are not stated in the announcement.
Do you manage Microsoft 365 in your organisation? Then record who may use agents and on what budget, before the bill becomes variable. You can find how Copilot scores on GDPR and price on the model page.
Google: Gemini 3.8 Live with a speaking avatar
Google has made Gemini 3.8 Live with Live Avatar generally available to business customers. The model conducts real-time voice and video conversations through an animated face, in 97 languages. According to Google, there are US and European endpoints. All audio and video carries an invisible SynthID watermark. Creating an avatar based on your own footage is only possible with Google's approval.
If you deploy something like this for customer service or marketing, your conversation partner must know they are speaking with AI. That follows from the transparency obligation in Article 50 of the EU AI Act. A watermark alone is not visible to your customer, so simply say so at the start of the conversation. You can read how Gemini handles your data on the model page.
What you can do now
- Check that training is switched off in every ChatGPT account in your team.
- Ask your Microsoft administrator how the use of Copilot agents is restricted.
- Compare at /modellen/ which providers keep your data out of training.