Reading Room

Outside your usual reading: this is a mobile security write-up, not an AI story, but the lesson travels.

A single text message can crash an Android phone's call app and drop any call in progress, including a call to emergency services, with no action needed from the phone's owner. Independent researchers reproduced the attack with a 100 percent success rate on a Pixel 9 running April 2026 software and on several Samsung Galaxy and A-series phones.

The message is a type called a WAP Push, a short format phones use to receive links and settings pushed from a carrier. Android's phone app reads a length value out of that message and is supposed to check that the value makes sense before using it. It does not. The flaw sits in a function called decodeWapPdu, inside a file named WapPushOverSms.java, where the code reads the length as a large number type, converts it down to a smaller number type with no range check, then uses that converted number as the size of a new block of memory to allocate. A crafted value decodes to a number that, once converted down, becomes negative 2,147,483,648. Asking the phone to allocate a block of memory of a negative size throws an error the phone app never catches, and the app crashes and restarts. The message never appears in the phone's inbox, and the owner never sees who sent it or that anything happened.

A single message produces one crash and one restart, a short, repeating outage each time it lands. Sending about 30 of the messages in quick succession is worse: the repeated crashes collide with the phone app's own writes to its message database at the same moment, and that collision corrupts the database file. Once that file is damaged, the phone app no longer needs a new malicious message to fail. It crashes every time the phone starts up, simply trying to read its own damaged records.

The researchers argue this is not a minor bug because of where it lands: a phone that cannot reliably place a call cannot reliably reach emergency services either. Losing emergency call access is treated as a public safety failure on its own, and this bug can manufacture that exact failure remotely, with a text message the victim never sees. The rule was tightened in 2024 after some phones were left able to make ordinary calls but not emergency ones during an older network shutdown, the same end result this bug produces by a different route.

The fix is small: add a check that rejects the length value before it gets converted and used, instead of trusting it. The researchers handed Google that exact fix the day they filed their first report. Google's own bug bounty program reviewed the report three separate times, including two appeals, and each time classified it as not a security issue. The program has its own category for exactly this shape of bug, a destructive remote denial of service, meaning it disables a device in a way that needs more than turning it off and on again to fix. The researchers published the full writeup and a working proof-of-concept only after the standard disclosure window had passed with nothing shipped.

The detail to keep past the specific Android bug is the shape of the failure: a boring, easily reviewed mistake, a number not range-checked before a type conversion, turned into a remote, no-interaction denial of service that can repeatedly disable a phone's calls, including emergency calls. That exact category of mistake is what code review and tests are supposed to catch before anything ships, in his own agents' output as much as in a phone's operating system.

A news roundup pulls four takeaways from Anthropic's filing for what could be a record breaking stock market listing, though the specifics aren't given in this save. Anthropic is the company behind Claude, the tool his whole operation runs on, so the terms of its listing are worth tracking as more detail surfaces.

Anthropic's chief executive is publicly calling for stronger government regulation of AI, according to an ABC News exclusive. Regulation sets the rules every AI company, including Anthropic, has to ship under, which shapes the tools his whole operation depends on.

OpenAI and ServiceNow are deepening their partnership: OpenAI's models and agents, meaning AI that takes several steps on its own, are going into ServiceNow's workflows, and ServiceNow's workflows are going into ChatGPT and Codex. It shows how fast AI vendors are wiring themselves into everyday business software, worth watching if he ever sells into companies running these tools.

Elon Musk made a short AI-generated film based on The Odyssey, the ancient Greek story, and said he wanted to beat a version made by director Christopher Nolan. Viewers who watched it called it low quality and far from the original story, a real data point on where AI video still falls short of a human-made production.

Oura pulled its public stock offering, but an investor who read the filing line by line found a different explanation: of 50 million shares on offer, 36.5 million were existing shareholders cashing out, leaving Oura only about $567 million against roughly $1.5 billion going to sellers, with about 93 percent of that already earmarked for employee tax bills.

Valuation professor Aswath Damodaran has published his own independent estimate of what Anthropic is worth, separate from whatever number its stock market listing lands on. A credible outside check on whether Anthropic's eventual public price reflects the real business or just IPO excitement.

Anthropic has launched new tools that show companies exactly what their AI spending is buying them. Spending visibility is the first thing any company, including his own ventures, needs before it can judge whether AI tools are worth the cost.

Starling Bank, a British challenger bank, says it is building what it calls the first banking AI model trained entirely on UK sovereign cloud, meaning cloud computing kept physically and legally inside one country rather than run by a foreign company. A newsletter that read past the announcement found the claim only holds with three qualifiers it did not spell out in full, plus a more basic problem: when Starling made the announcement, the model had not actually been trained on real customer data yet.

The model is due to start running Starling's fraud detection from early 2027, and later power an AI assistant for the bank. That timing matters because at least three other lenders are already running similar transaction-monitoring AI models today, and one of them is already doing it on a non-US cloud, which cuts against how novel Starling's first claim really is.

The sovereignty claim has a hole in it. Starling's sovereign setup routes customer data through Google Cloud before that data ever reaches Civo, the UK-owned cloud provider that is supposed to be the sovereign part of the story. So the data still passes through a large American cloud company on its way to the British one, even as the bank markets the setup as independent of big technology infrastructure.

The newsletter's fuller comparison, including who built similar banking models first and with how much more training data, sits behind its paywall, so that detail is not available here. What is confirmed is the pattern worth watching: a sovereign AI label applied to a system that, underneath, still depends on the same handful of large cloud and chip companies everyone else depends on. For anyone evaluating a vendor's claim of independence from big technology, the Starling case is a reminder to ask what actually touches the data before it reaches the sovereign part of the pipeline, not just what the marketing says.