Some bugs are ignored because nobody can exploit them, until someone can. Today’s briefing starts with that problem in PHP, then moves through a 60,000-person Microsoft buyout offer and a Meta defector with a very small team and a very large target.
1. Bugs nobody could exploit are now being exploited
The PHP flaws known as PHP2Shell sat unpatched for two years because the foundation behind PHP judged there was no practical way to exploit them. CISA now says attackers are doing exactly that, and has given civilian federal agencies three days to fix their systems.
According to TechCrunch, researchers at Calif found a way to exploit the dormant bugs earlier this year with help from an AI model. Patches shipped less than a week ago, and some governments and large companies took sites offline to apply them. PHP is reportedly used on about 70% of web-facing servers. CISA didn’t say who is exploiting the flaws. The Register’s headline says researchers have clocked successful exploits, though the text we have is thin.
My read: the lesson is less about patch speed than about triage. “Hard to exploit, so low priority” is a weaker judgment when a model can help find a working path, including by chaining flaws. Underfunded open-source projects can’t easily buy the security reviews that would catch these earlier. The open question is how many other shelved bugs are waiting for the same treatment.
2. A Meta AI director quits to build Office without the black box
Rohit Patel spent eight years on AI at Meta. He has now left to start Ndoc, which he pitches as an AI-native replacement for Office or Workspace. The most interesting part of the pitch isn’t the AI. It’s that the AI shows its work.
Ndoc puts slides, written analysis and underlying Python scripts in one workspace, so a user can ask the chat what a script does and check the formula. Patel also plans to open source a single-file “.ndoc.html” format. He says doing that at Meta would have meant input from “hundreds of people.” That is a more specific reason to leave than any broad claim that Meta’s AI bets are failing, and TechCrunch’s reporting doesn’t support the bigger version.
The odds are long. Ndoc has three people and no funding, and it is in alpha with a full launch planned for next year. OpenAI, Anthropic, Google and Meta all sell enterprise AI tools, and OpenAI’s Space is a more direct competitor. Patel says early users cared little about his lower running costs and more about having everything in one place. A Register headline today puts the share of boardroom execs seeing AI savings at 3%, but we only have the headline. If buyers are still skeptical, auditability may be the stronger pitch.
3. Opera’s WebAI shifts AI costs from developers to users’ laptops
Opera’s pitch for WebAI is that developers can add AI features without running AI infrastructure or paying for every request. The catch is that the bill doesn’t vanish. It lands on the user’s hardware.
TechCrunch reports that WebAI lets pages run open-weight models such as Google’s Gemma directly in the browser. On Macs it also supports Apple’s on-device model. The feature is an opt-in early preview in Opera One, with Opera GX and Opera Neon to follow. Opera says users keep their data on their own devices and need no AI subscription.
The same article notes the trade-off: local models demand more from the machine, and performance depends on the hardware. That matters for developers. A feature that works well on a new laptop and crawls on an old one is harder to promise than an API call. Every source here is Opera’s own announcement, and only Opera browsers are named. Whether this becomes a real web standard depends on whether anyone else adopts it.
4. UK ministers found time for datacenter bosses, not environmental groups
The Register reports that an investigation found businesses dominated government meetings on AI Growth Zones, while communities were left out of decisions. The headline says UK ministers had plenty of time for AI datacenter bosses but none for environmental groups.
We only have the headline and standfirst, so the numbers and methodology are missing. The finding is still a useful signal for anyone planning UK infrastructure around these zones. The people shaping the policy appear to be the ones who will build in them.
That cuts both ways. A friendly process is good for operators. But decisions made without local input can invite objections later, and the evidence here doesn’t show whether that will happen.
5. Microsoft wants 60,000 people to leave on their own
Microsoft is offering voluntary buyouts to about 60,000 employees, which Ars Technica describes as a quarter of its workforce. The Register says the offer comes with an end-of-year deadline. Ars reports that staff in the US, Canada, India and the UK may receive offers today.
The supplied sources don’t say why, so any link to Microsoft’s AI spending is inference. What the format does tell us is that this is voluntary. A company that lets people choose to go gives up some control over who leaves, since the people with the best outside options are often the first to take a check.
That’s why the details matter. We don’t yet know how many accept, how generous the packages are, or whether Microsoft follows with involuntary cuts. Those details would show whether this is a cost cut or a reshuffle.