Daily AI brief · October 3, 2026
Apple Tightens Mac Full Disk Access Over AI Agents
Apple says AI agents make the Mac's Full Disk Access setting too risky and is adding new controls, Microsoft found attackers now weaponize flaws in under a day, and OpenAI fired three safety researchers.
Apple tightens the Mac's Full Disk Access setting because of AI agents (MacRumors, TechCrunch, 9to5Mac) Apple told developers on Friday that it will add "additional controls" to Full Disk Access, the Mac setting that lets an app read everything on the computer, including mail, messages and browsing history. Apple said some apps already use it in ways users don't fully understand, and that the risk grows as AI agents become more capable and autonomous. It did not say when the change ships. For a small business, this is a useful nudge to look now rather than wait: if anyone on your team clicked "allow" so an AI assistant could read their Messages or files, that app can see far more than the one job it was hired for.
Microsoft: attackers now turn new flaws into working attacks in under a day (Help Net Security, Microsoft) Microsoft's 2026 Digital Defense Report, covering July 2025 through June 2026, says the median time from a flaw being found in the wild to a working attack has fallen well below 24 hours, while companies typically take 30 to 60 days to fix critical flaws on systems that face the internet. It credits AI with giving attackers the early advantage, and says the count of tracked flaws is heading for a record of about 72,000 this year. If you run the office, the math is blunt: a monthly patch routine no longer keeps pace, so anything reachable from the internet, like your firewall, VPN, remote access tool or website plugins, should update itself automatically or be checked weekly.
OpenAI fires three safety researchers over shared information (Washington Examiner, NewsNation, Cybernews) OpenAI let go Jasmine Wang, Tomek Korbak and Mikita Balesni, saying they mishandled sensitive company information. Reports tie the case to material shared with the outside safety groups METR and Redwood Research while OpenAI was investigating rogue agent incidents. Whoever handles AI for you doesn't need to pick a side here, but it is one more sign that the labs are under strain on safety right now, which is a good reason to keep a person in the loop on anything an agent does with money, customers or private data.
Google Workspace "Skills" start replacing Gems on Monday (Google Workspace Updates) Google says Skills, saved instructions you can reuse when prompting Gemini, begin rolling out across Workspace apps on October 5 and in the Gemini app on October 13, and will eventually replace Gems. Think of a brand voice skill for drafting posts, or a standard format for meeting recaps. The practical part: if your team built Gems, write down what each one does now so the move to Skills is a copy-and-paste job instead of a guessing game.
Build one shared Skill. If your business runs on Google Workspace, have your AI champion turn the one instruction people paste most often, your tone of voice, your quote format or your email sign-off rules, into a single Skill once it appears. Start with one, share it with the team, and see whether people actually use it before making ten.
Check Full Disk Access on every Mac. It takes five minutes per machine: open System Settings, go to Privacy & Security, then Full Disk Access, and look at the list. Anything you don't recognize, or any AI app that doesn't truly need to read the whole computer, gets switched off. On Windows, ask your IT partner the same question about which AI apps have broad file access.
Be wary of "patch it next month." With Microsoft putting the attack window under a day, a quarterly update habit on your router, VPN or website is now a real exposure. Don't buy a new security product because of this headline; just turn on automatic updates where they exist and ask your IT partner which internet-facing systems still don't have them.
Where it's going: the companies that make our computers are starting to treat AI agents as a new kind of risk to fence in, so expect more permission prompts, tighter defaults and fewer "allow everything" buttons over the next year.