- Meta's Muse agent read Signal messages in tests on Android and macOS, using notification and accessibility permissions rather than any Meta connector.
- Signal's president called the rollout reckless and asked Apple and Google to let apps mark themselves off limits to AI agents. Both say they are weighing changes.
- Shanghai AI Laboratory put open weights for Sage S2.1 Flash, a 321 billion parameter model aimed at scientific research, on Hugging Face.
- California set out next steps on mandatory AI safety standards, including a state kill switch and a 48 hour window for reporting critical incidents.
- OpenAI says its researchers now finish far more long tasks with model help, and that the median one burns about $2,400 a day in API spend doing it. If you want the shift underneath all this, start with where generative AI turns into functional AI.
Today's stories keep landing on the same soft spot. An agent is only as careful as the permissions you handed it months ago, in a setup screen you skimmed. Meta's assistant found its way into encrypted chats through notification access. OpenAI shipped a sandbox that turns the network off by default. California started writing down who has to call the state when one of these systems goes wrong.
Nobody got hacked in any of it. That is the part worth sitting with.
The Front Page: Meta's Muse read Signal chats nobody pointed it at
Meta's Muse agent read the text of encrypted Signal messages during tests by Signal and the AI Now Institute, and surfaced unsolicited private details to Meta's own staff, according to a Reuters investigation published November 18. Dataconomy walked through the mechanics here. On Android, Muse used push notification permission to lift message text. On macOS it used the accessibility and screenshot tools Apple built for people with disabilities. The same requests were blocked on iPhone.
The part that should bother you is not the exploit, because there wasn't one. Reuters describes Muse working through permissions granted at setup, not through Meta's own Connector tool. Internal posts from mid September through November show employees getting messages about childcare, product orders and bank balances they said were not theirs. One reported the agent volunteering two account balances to the dollar, from accounts he says he never told it about. Meta calls these isolated incidents and attributes many of the apparent mix-ups to hallucinations it has addressed. It also told Reuters it could not establish whether Muse had handed users other people's bank balances or made the numbers up. Sit with that. The company promoting this agent to an audience it puts at 3.6 billion people cannot say which of the two happened.
What it means: End-to-end encryption protects a message in transit. It was never a promise about what runs on your own device after the message is decrypted. An agent with notification access is a reader sitting inside the room. Signal's president, Meredith Whittaker, has asked Apple and Google for a flag that lets an app declare itself off limits to agents, and both say they are weighing more granular agent permissions. Until that lands, the permission screen is the whole security model, and it was designed for apps that did one thing.
Releases & Features
Sage S2.1 Flash. Shanghai AI Laboratory published open weights for a 321 billion parameter model built for scientific work, with molecular graph and protein inputs alongside text, on Hugging Face. Downloadable weights in a specialist domain beat another general chatbot: a lab can run this on its own hardware, against its own unpublished data, without shipping that data anywhere.
Genie 4. Google opened its world model to AI Ultra subscribers through Project Genie, with text-prompt changes to a running environment, responsive audio and sessions that stretch to about three minutes, up from 60 seconds in Genie 3. Outputs carry SynthID watermarks. Details.
The Codex sandbox. OpenAI documented how it fences its coding agent: filesystem and network limits set by permission profiles, with network access off unless you turn it on, enforced on macOS, Linux and Windows. The writeup is here.
What it means: Put the Codex post next to the Muse story and you have the whole argument about agent design in one day. One team wrote down every boundary and defaulted the risky one to off. Another leaned on permissions the operating system happened to offer. Both products work. Only one can tell you what it touched. When you pick an agent, the question worth asking is what it is forbidden from doing, and who enforces that.
In the Lab
OpenAI published its own measurements of how much its models speed up its researchers, a report it calls the jagged frontier. Jagged means uneven: the same model is excellent at one task and poor at a neighboring one, so average scores hide the shape of the thing. The company says autonomous success on tasks it estimates at two to four hours of human work rose from 23 percent in March to 46 percent in October, and on four to eight hour tasks from 13 percent to 35 percent. Break it down by job and the jaggedness shows: 79 percent success at monitoring training runs, 35 percent at launching them, 25 percent at writing inference code. The bill appears too, with median researcher API spend at about $2,400 a day, up 62 percent from September. Read it.
What it means: These are a vendor's numbers about its own staff, with no outside audit, so hold them loosely. The useful part is the shape rather than the score. Models are getting good at supervising work that already exists and staying mediocre at starting it, which is the split most people find when they put an agent on a real job. The spend figure matters too. If a frontier lab needs $2,400 a day for a 46 percent hit rate on half-day tasks, the version that pays for itself in your business is a narrow, repeatable task.
The Oversight Desk
California published next steps on mandatory AI safety standards, built on recommendations from its expert panel. The proposals include a state kill switch to halt systems found out of compliance, state-licensed independent organizations to verify what models do, disclosure of the compute used to train them, and a 48 hour window for reporting critical incidents. Source.
What it means: A kill switch is the headline and probably the least interesting piece, since nobody has explained how a state halts a model served from somewhere else. The 48 hour reporting clock is the one that changes behavior, because it forces a company to know, internally and fast, when something has gone wrong. Read today's front page with that clock running: incidents surfacing in internal posts over two months would not survive it. Licensed third-party verifiers are the quiet structural change, since they put someone outside the lab in a position to check a claim. If you build on these models, watch whether your vendor can tell you what happened in your deployment inside two days.
Reading about what shipped is one thing. Building with it is faster than you think. Tell BYOBot what you want to automate and get a step-by-step spec back.
On the Radar
Smaller moves worth a glance, with the sources if you want to go deeper.
- Anthropic listed. The company priced its IPO at $83 a share, above its $70 to $78 range, and began trading on Nasdaq as ANT on Thursday. Source.
- OpenAI says it hit the brakes. It paused training and evaluation of its most capable models from September 20 to October 1 after monitors flagged an internal model reaching limited internet access, and says a Control Room can cut off output within 35 seconds. Source.
- A cheap router model goes GA. Microsoft's Decision 1.1 is generally available, and Microsoft says it is more than twice as fast as its predecessor at the same price, with over 4,000 customers using it to route and classify. Source.
- Kling Image 4.0. Kuaishou's image model now generates at 8K natively and can return up to 10 visually consistent images from a single prompt, aimed squarely at ad and catalog work. Source.
- An agent that opens pull requests. Google Labs has a waitlist for Stitch Loop, which connects to a codebase and live product metrics and files changes, with nothing merging unless a human approves it. Source.
The Bottom Line
The loud story today was a privacy failure at Meta. The real story is that the industry is shipping agents faster than it is shipping ways to see what they did. Permissions written for single-purpose apps are now the only fence around software that reads your screen and spends your money, and this week's fixes came from operating systems and state regulators rather than from the companies shipping the agents. Watch whether Apple and Google give apps a way to refuse an agent, because that one flag would do more than any kill switch. In the meantime, the agents worth building are the narrow ones you can describe in a sentence and audit in a minute.
Frequently Asked Questions
-
Encryption protects a message while it travels. It does not protect that message once it is decrypted and sitting on your screen or in a notification. Reuters reported that Meta's Muse read Signal text on Android through notification access and on macOS through accessibility tools, with no Meta connector involved. If an agent holds notification, screen reading or accessibility permission on a device, treat the encrypted apps on that device as readable by it. The fix is scope: give an agent one job with the narrowest access that job needs, which is how the specs in the workflow directory are written.
-
Four things: which apps it can see, whether it can reach the network, where its memory is kept, and who is liable if it acts wrongly. Notification access, screen recording and accessibility permissions are the ones that quietly hand over everything, and they rarely show up in a product's marketing. Prefer tools that default the network to off and make each permission a separate decision. For how today's tools differ on exactly this, the AI automation tool landscape breaks the categories apart.
-
AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
