September 23, 2026 — The AI Price War Just Got a Security Problem

Host0:39

Anthropic put out Claude Opus five point five yesterday, and it did something a little unusual for a flagship model. It got cheaper. Anthropic says Opus five point five performs about as well as its own Fable five point one on most tasks, but costs forty percent less to run than the old Opus five. Input tokens run four dollars per million, output tokens twenty dollars per million, and cache reads drop to twenty cents per million. Anthropic also says it generates responses more than thirty percent faster than its predecessor. One early tester quoted on Anthropic's own product page said, in effect, that the model writes the way they do, which is a nice compliment until you remember Anthropic wrote the page. The company also says Opus five point five ships with what it calls preserved thinking, an anti-distillation safeguard meant to stop rivals from cheaply copying the model's reasoning. It was tested before release by outside evaluators including Frontier Design and METR. Here's the thing. Within hours, OpenAI answered. It rolled out two updated members of the GPT six family, called Sol and Luna, and according to OpenAI's own announcement as reported by TechCrunch, they'll cost roughly half what the five point six versions cost. OpenAI also claims Sol makes about half as many mistakes as its predecessor on the company's internal factuality test, reaching what it calls Astra level reliability at a much lower price. Notice the pattern. Anthropic's numbers come from Anthropic. OpenAI's numbers come from OpenAI. Nobody outside either building has run these claims yet. Independent benchmarkers will be the ones to check whether the accuracy gains are real or just an artifact of different default settings and caching tricks. Until that happens, the open question is simple: are you getting a better model, or a cheaper way to call the same one. Cheaper models on the cloud side is one way to lower the cost of running AI. The other way is to stop sending the work to the cloud at all, which is exactly what Qualcomm is betting on.

2:58

Qualcomm announced two new phone chips yesterday, and the names alone tell you the strategy: the Snapdragon eight Elite Gen six, and above that, the Snapdragon eight Elite Extreme Gen six. Both are built on a two nanometer process, with Oryon CPU cores Qualcomm says can hit peak clock speeds of five gigahertz. The Extreme version, aimed at the very top tier of phones, reportedly gets around fifty percent more shared memory than the standard chip, along with upgrades to the Hexagon neural processor and the Adreno graphics core. Translate that out of spec-sheet language. More memory on the chip means the phone can hold a bigger AI model in its own head instead of constantly phoning a data center. That's the whole point of what Qualcomm calls the agentic AI age. Early reporting ties the Extreme chip to the Motorola Signature twenty seven, expected to be one of the first phones out the door with it. The catch is that a chip on a spec sheet and a chip in your pocket that doesn't overheat mid-scroll are two different products. How this holds up under real thermal conditions, and how much OEMs charge for phones with the Extreme variant, is still unknown. Cloud providers, who'd rather keep that inference workload paying rent on their own servers, have their own reasons to doubt this moves much work on-device. More powerful chips running more autonomous software on more devices is great, right up until the software running on them isn't yours. Which is a decent description of what Microsoft spent yesterday cleaning up.

4:39

Picture a criminal subscription service, sold on Telegram for fifteen hundred dollars to join and five hundred dollars a month after that, and the pitch is basically: we'll have an AI read the stolen inbox for you. That's what Microsoft says it dismantled yesterday, a platform it calls EvilTokens. According to Microsoft, as reported by Ars Technica, the service used an AI-style chatbot to scan a compromised inbox and flag the good stuff: trusted relationships, payment authorizations, anyone with sensitive responsibilities. A target list, generated automatically. Microsoft says EvilTokens users compromised around twelve thousand customer accounts across roughly ten thousand organizations worldwide. The technical trick was abusing device code authentication inside Microsoft's own identity system, Entra, to sneak an attacker's device onto a victim's account without ever needing the password again. Microsoft, working with partners including SpyCloud, seized fifty websites and one hundred fifty domains tied to the operation. The UK's Metropolitan Police arrested two men in connection with it. Microsoft's own warning is worth repeating plainly: once an inbox is compromised, assume criminals understand its contents in minutes, not days. Worth noting, that twelve thousand figure is Microsoft's own operational estimate, not an independently audited count, so treat it as a floor rather than a final number. Here's the stranger cousin of that story. Cisco's Talos research team published findings on malware called CLOSEDQUORUM that doesn't have one command server. It has four. The malware polls up to four commercial AI providers, Talos names DeepSeek, Qwen, Mistral, and Google's Gemini, asks each what to do next, and goes with whichever answer wins a majority vote. It's a sixteen point four megabyte, sixty four bit Windows program written in Go. Talos is careful to say the public copy they found has placeholder API keys and a fake webhook, and they have no confirmation it's been used in a real attack. But the code is linked to forum posts from a developer going back to twenty twenty-five. Talos calls this effort displacement, handing the decision-making phase of an attack over to the AI itself, so the operation can keep running without a human clicking anything. Whether that's a live threat or a lab experiment, the question for the AI companies whose products got named is the same: can you spot your own API being polled for a malware vote, and how fast can you cut it off. Microsoft and Talos are describing criminals borrowing AI. The next story is a hacking group claiming they didn't need to borrow anything, they just walked straight into a federal agency's server room.

7:45

A group called ShinyHunters posted on its leak site yesterday that it had broken into the FBI and taken data on, in their words, almost all FBI agents and everyone who ever applied for a job there. TechCrunch reviewed the leak site post. Four zero four Media went further and checked a sample of the stolen names, home addresses, and phone numbers against public records, and said the sample checked out. The group claims it took terabytes of data total. According to the hackers' own account, as reported by TechCrunch, they say they got in through an Oracle PeopleSoft server and then pivoted into a government cloud environment hosted on Amazon. They also claim to have defaced the FBI's jobs website. And in a detail that's almost stranger than the breach itself, ShinyHunters says the hack wasn't financially motivated at all. They say they want the FBI to take down a report the group claims contains false allegations against them. As of that report, the FBI hadn't responded to requests for comment. So what you have right now is a claim, partially corroborated by two outlets checking a slice of the data, sitting next to total silence from the agency that would actually know. That gap, between what's been sample-verified and what's just asserted, is exactly where this story could go either way.

9:11

Four stories, thirty seconds each, starting now. AWS CloudWatch Omni. AWS announced CloudWatch Omni yesterday, a new observability tool built on OpenTelemetry with a dedicated login, no AWS console required, plus an AI agent that helps investigate problems in natural language. If your company is running AI agents in production, Amazon just built the dashboard to watch them, which is either reassuring or exactly the lock-in a rival cloud provider will warn you about. Snorkel AI. Snorkel AI raised three hundred fifty million dollars in a Series E, tripling its valuation to three point five billion dollars. The company makes training data for frontier AI labs, and its CEO is calling it the frontier lab for agentic data, though how much of that revenue stays with Snorkel versus passing straight through to human labelers is still an open question. Cyera. Data security company Cyera added four hundred million dollars to its Series G from Goldman Sachs Alternatives, on top of a June round that valued Cyera at twelve billion dollars. The money's going toward tools named Agent Guardian and Cyera Endpoint, built to police what AI agents are allowed to touch inside a company, which tells you agent security is about to become as big a headache as the agents themselves. Apple. Apple released the first public betas of iOS twenty seven point two and macOS twenty seven point two yesterday, headlined by a redesigned Health app with new Insights and Longevity tabs, including a feature called Health Age that estimates your fitness age against your actual age. So now your phone can tell you you're doing worse than your birth certificate suggests, which is a bold new way to start the morning. That's Fast Boot. Four stories, zero mercy.

11:12

Nat Friedman, who runs product for Meta's Superintelligence Labs, posted something on X yesterday that read less like a denial and more like a confession with good manners. Meta's new assistant, Muse, had been getting compared online to an open-source AI agent framework called OpenClaw, after users noticed Muse contained files with names like SOUL dot M D that looked nearly identical to files inside OpenClaw. According to TechCrunch's reporting on Friedman's post, he said Muse was, in his words, definitely heavily inspired as a product by OpenClaw, while insisting Muse was built from scratch. He went further, praising OpenClaw's creator, Peter Steinberger, saying the team thought Peter got those things exactly right. Meta's spokesperson, when TechCrunch asked for more, just pointed back to Friedman's post and said the company had nothing further to add. Which leaves you in a strange middle ground. Heavily inspired and built from scratch aren't necessarily contradictions, but near-identical file names make that a harder sentence to say with a straight face. Nobody's filed anything yet. But the moment a company that size admits to copying another product's structure this closely, the question of whether any of that content carries licensing obligations stops being hypothetical.

12:39

So that's the day. Anthropic and OpenAI both cut prices on their top models within hours of each other, Qualcomm built phone chips for AI that stays on the device, Microsoft dismantled an AI-assisted fraud ring while Talos flagged malware that lets AI chatbots vote on its next move, ShinyHunters claims it walked out of the FBI with agent and applicant data, and Meta admitted its new assistant borrowed more from OpenClaw than it first let on. Full sourcing for every story is in the show notes on the website. If this show's useful to you, follow The Daily Reboot wherever you're listening, so tomorrow's episode just shows up. That's the show. Talk to you tomorrow.