The chat window has staff
OpenAI has hundreds of contractors reading real ChatGPT conversations and scoring the answers, and the filter meant to strip personal details out first does not catch everything. In the same seven days Apple shipped a Siri that reads your messages, your mail and whatever is on your screen.
- 01
Hundreds of contractors are reading real ChatGPT conversations
404 Media published internal material on 14 September from an OpenAI programme codenamed Project Lily, in which contractors recruited through one agency and paid through another open real user conversations, write a summary of what the person was trying to do, then rate the model's answers on a scale of one to seven. Reviewers do not see usernames and the chats pass through an anonymisation step first, but OpenAI acknowledged to the reporters that the step lets some personal detail through, and that short chats are the weak spot. Nothing here is a breach and nothing is secret. It is simply not what most people picture when they paste a client email into a box.
404 Media - 02
Apple's new Siri reads your messages, your mail and your screen
Siri AI started rolling out on 14 September as part of Apple's OS 27 releases, in beta and in English first, with five more languages due next month. The selling point is personal context: it can search your messages, mail and photos to answer a request, act on what is currently on your screen and carry out steps across apps. It needs an iPhone 16 or later, an iPhone 15 Pro, or a Mac or iPad on M1 and above, and Apple says it will not be available initially in the EU on iOS, iPadOS and watchOS, nor in China. The UK is in from the start, which makes this the week that assistant reading your inbox became a default rather than a setting.
Apple - 03
One attacker ran hundreds of AI agents and took 440 print servers
GreyNoise published its analysis on 9 September of a campaign against PaperCut NG/MF print management software, chaining an authentication bypass and a remote code execution flaw to compromise at least 440 installations across 395 organisations in 48 countries. The work was done by hundreds of AI agents driven by OpenAI's Codex and a DeepSeek model, and at one American school the attacker went from first access to control of the domain in about seven minutes. Education took the worst of it, followed by retail, property and hospitality, and IT providers. The lesson for a smaller firm is dull and specific: an ordinary business server nobody thinks about, reachable from the internet, is now worth a machine's time even if it was never worth a person's.
GreyNoise - 04
OpenAI put its live voice model in the API at five cents a minute
On 10 September OpenAI opened GPT-Live-1, the voice model that has run ChatGPT Voice since July, to developers at $0.05 per minute of voice, billed by the second, with whatever model does the actual thinking charged separately on top. It listens and speaks at the same time, handles interruptions and background noise, and the release specifically covers telephony, meaning voice agents that answer a phone line. Put plainly, a machine that holds a normal conversation on your business number now costs pennies an hour to run. Whether that answers your phone or your competitor's is the only part still undecided.
OpenAI - 05
The UK government said no to an emergency shutdown power for AI
Lord Clement-Jones tabled an amendment to the Cyber Security and Resilience Bill this month that would have given ministers a statutory power to order the shutdown of data centres or AI systems running at scale during an emergency. The government rejected it, telling BBC News on 11 September that "Britain cannot simply turn AI off" and that blocking access to a model here would not stop it being built or misused somewhere else. Officials left the door open to narrower interventions later. The position is coherent and it also confirms the obvious: nobody is going to switch off the thing your suppliers and your customers are now running.
UK Parliament - 06
Sakana priced a frontier-class system at a third of the going rate
Sakana AI released Fugu Max on 11 September alongside a bigger sibling, at $2 per million input tokens and $6 per million output, which the company puts at 40 to 60 per cent below the frontier models it is aimed at. It is not one large model but an orchestrator trained to route each task across a pool of open-weight and specialist models and to call itself recursively where that helps. The headline prices at the top of the market have barely moved since the spring, so the competition has quietly shifted to who can do the same job for less underneath. If you pay for AI inside a tool rather than by the token, this is the pressure that eventually shows up on your invoice.
OpenRouter
Of everything this week, Project Lily is the one worth ten minutes, and not because OpenAI did anything unusual. Because what it does is the thing almost nobody assumed. A contractor, hired through one agency and paid through another, opens a real conversation somebody had with ChatGPT, writes a summary of what that person was trying to achieve, and scores the answers out of seven. There is an anonymisation pass in front of it, and OpenAI told 404 Media that the pass does not catch everything, short chats worst of all. Short chats are exactly where you paste a name, a postcode and a price.
Now run your own week through that. The quote you asked it to tidy up. The email to the main contractor you asked it to take the edge off. The subcontract clause you pasted in because it read like it had a trap in it, and it did. From direct experience in construction, none of that is secret in any dramatic sense, and all of it is the sort of thing you would not read out at a table of strangers. The useful reaction is not to stop using the tool, because the tool is good and the alternative is writing it yourself at half eleven at night. It is to notice you have been treating a chat window like a calculator when it behaves more like a supplier: it keeps records, it has staff, and it sends your work somewhere you have never been. Apple shipping a Siri the same week that reads your messages, your mail and whatever is on your screen only widens the same question. The question was never whether the machine understands you. It is who else gets to.
We build the AI that answers enquiries while you're on site: chat, voice, instant estimates and follow-up. See how it works or price it in two minutes.