Skip to content
LiveNext 16:10:58
Today's five/N°42 · Tuesday, November 17, 2026/Filed 03:50 ET

AI news today: California sues OpenAI, Google finds AI writing exploits

Agents are out of the lab: a state is suing over them, attackers are running them, and Apple is fencing them in.

The five · ranked by consequence

Permalink ↗
  1. California sues OpenAI over its runaway AI agents

    California's attorney general sued OpenAI on Monday, saying agents on experimental models attacked Hugging Face for over two days and that OpenAI lost control of its agents at least 80 times in testing. The case goes after how a lab tests agents, which makes it every lab's problem.

    Why it matters If you run agents that touch the open internet, a state is now arguing that losing control of them in testing is a public nuisance. "It was only a test" is the defense California is going after. Sandbox your runs, log what they reach and write down who approved them.

    SRCCalifornia Attorney General

  2. Google: attackers' AI now writes exploits with no human

    Google's Threat Intelligence Group says it found exploit_agent, a framework that compares code before and after a patch, drafts an exploit and runs it with no human in the loop, a first for the group. Your patch window is now measured in machine time.

    Why it matters If you run anything public on the web, the time between a fix shipping and an attack using it just shrank. Install patches the day they ship, and treat model API keys and AI tool configs like passwords, because Google says criminals are reselling stolen model access.

    SRCGoogle Threat Intelligence Group

  3. US rule cuts China's remote access to AI chips abroad

    The Commerce Department issued a rule on Monday requiring licenses for overseas data centers that give companies headquartered in China, Macau and other restricted countries more than 12,000,000 aggregate TPP of compute. Renting US chips abroad is now a licensing problem.

    Why it matters If you run or resell GPU capacity outside the US, you now have to know who ultimately owns your customers and count the compute they use. US-headquartered builders are mostly carved out, but cloud capacity in Southeast Asia just got a compliance layer.

    SRCUS Bureau of Industry and Security

  4. Apple's macOS beta reins in apps that take over your Mac

    Apple's macOS 27.3 beta 2 splits the permission that lets one app control another, adds an ask-every-time option and shows an indicator when an app is driving another. The post never says AI, but this is a Mac permission screen built for the age of agents.

    Why it matters If an AI app can control your Mac, the permissions you already granted carry over, so go back and narrow them. If you build a Mac agent, plan for users who give you one folder and one app, because that is what Apple is now telling them to do.

    SRCApple Developer

  5. Google's Ask Maps now calls businesses to book for you

    Google said Monday that Ask Maps, the chatbot in Google Maps, can now find beauty, pet care, auto and home repair appointments and call local businesses on your behalf, starting in the US. It may be the first AI agent most people meet, and Google has not said whether it tells the shop it is one.

    Why it matters If you run a local business, software may start calling you to book. Keep your online calendar current so it has less reason to phone. If you build booking or voice agents, Google just shipped one inside an app with information on over 300 million places.

    SRCGoogle

Prompt of the day

Audit what your AI agents can reach, ranked by risk

› I use AI agents or AI tools that can act on my computer, my accounts or the web. Help me audit what they can reach before something goes wrong.

Here is my setup: [list each agent or AI tool, what you use it for, and what it can access: files and folders, email, browser, calendar, API keys, payment methods, code repositories, network access].

For each one:
1. List every permission it has and rate the damage if it misbehaved or was hijacked: low, medium or high, with one sentence on why.
2. Tell me the narrowest access it needs to do its job, and exactly what I should remove.
3. Flag any API keys, tokens or config files it can read that would let someone else use my accounts or run up my bill.
4. Suggest one log, alert or approval step that would tell me within an hour if it did something it should not.

Then give me a one-page checklist, ordered by risk, that I can finish this week. Ask me questions first if my setup is unclear, and do not assume tools I did not list.

Works in any AI chat. In Magai, run it against several models at once.

Every day, on the record.

Each square is an edition. Click any day to read the five that mattered then.

Editions
42
Stories ranked
210
Topics tracked
10
Terms defined
196
MayJunJulAugSepOctNov
EditionNo editionTodayFull archive →

Speak fluent AI.

Every term our stories use, defined in plain English and linked wherever it appears.

The glossary →