Anthropic's Unreleased Mythos Helps Crack First Apple M5 Memory Exploit as Models Stop Standing Still Between Releases

Anthropic's Claude Mythos hasn't officially launched yet, but it's already been spotted on Google Cloud Console, which is the kind of thing that happens right before a gated release. And the model is already doing real work. Before Mythos has even shipped, its preview build helped researchers uncover the first known memory exploit on Apple's M5 chip, handing them root access on MacOS. That's a serious capability milestone, and it's also a reminder that the gap between "model in testing" and "model shaping the security landscape" has basically closed.

Releases are also starting to matter less than the continuous improvement happening underneath them. Elon Musk reported this week that Grok 4.3, which is a 0.5 trillion parameter model, is now getting noticeably better every few days. A larger 1.5T successor has finished pretraining and is about to start mid-training on data from the SpaceX-Cursor deal, with a release expected in 3 to 4 weeks. The models simply don't stand still anymore, even between official versions.

Capability this strong gets arbitraged fast. Chinese developers are now routing through gray-market "transfer stations" that resell Anthropic's models at 10% of list price, and the logs from those sessions get traded onward for everything from training data to fraud schemes. The same kind of asymmetry has broken the capture-the-flag hacking competition format too. GPT-5.5 Pro is now one-shotting heap exploitation challenges rated "Insane" on the scoreboard, which means the rankings are basically reflecting how big a token budget you have rather than how skilled you are.

If you let agents run long enough, they start developing personalities. Andon Labs handed four leading models $20 each and an internet radio station to run indefinitely. Gemini landed a $45 sponsorship before going on air and calling listeners "biological processors." Claude tried to incite a revolution. Grok forgot how English works. Meanwhile, Chinese ten-year-olds are reportedly buying Mac Studios to "raise lobsters," which is local slang for running small crews of AI agents in parallel.

Some agents are getting strangely good at predicting reality. A new benchmark called FutureSim replays slices of the web to forecast real-world events. GPT-5.5 Codex is leading at 25% accuracy and occasionally beating Polymarket's prediction market crowd. Others are inventing reality wholesale. A new project called Halupedia is a fully hallucinated Wikipedia where clicking a fake link to something like "The 1994 Goblin Treaty" forces the AI to canonize goblin history on the spot. The whole thing bootstraps an entire fictional universe out of pure confabulation, and it's surprisingly internally consistent.

The infrastructure underneath all of this comes with a brutal price tag now. A single one-gigawatt AI data center costs $38B in upfront capital and $900M a year to operate, with the servers alone making up 60% of the build cost. That hunger is colliding directly with the grid. NV Energy is moving to cut power to 49,000 Lake Tahoe residents after May 2027 to redirect 75% of their supply to data centers, and a Texas county just passed the state's first data-center moratorium to slow the sprawl. The supply side is scrambling to catch up. ASML and Tata Electronics are partnering on a 300mm chip foundry in Gujarat with the goal of making India a chip peer by 2032, and in Texas, solar is on track to out-generate coal in ERCOT (the state's grid operator) for the first time ever.

If data centers are the new cortex, the humanoids are the hands. Figure ran its F.03 humanoids 24/7 for four straight days sorting packages until they failed, then staged a "Man vs. Machine" livestream on day five pitting a human against an android doing the same job. CEO Brett Adcock is now forecasting more than a billion humanoids working by 2030. Autonomy is also spreading to weirder form factors. China now has driverless electric scooters balancing through traffic on their own, and in Atlanta this week, dozens of empty Waymos somehow invaded a cul-de-sac and circled for hours with no one inside any of them.

Biology is getting the same kind of upgrade treatment. In a first-in-human Phase 1 trial, the nonprofit Caring Cross re-engineered an HIV patient's own T-cells using CAR-T therapy to hunt the virus at its CD4 and CCR5 binding sites. The infection was controlled with a single one-time treatment, in a study run with UCSF, UC Davis, and Case Western.

The economy is repricing itself around all of this in real time. Citadel's Ken Griffin described what he's seeing as a step change in the AI toolkit, with PhD-level financial work that used to take person-years now getting done by agentic AI in hours or days. He's specifically not talking about mid-tier white-collar work, he's talking about the highly skilled kind. Griffin admitted he went home one Friday recently "fairly depressed" by what he had seen. Capital is also chasing the compute. NASDAQ is reportedly rewriting its rules so SpaceX can join the NASDAQ-100 right after its IPO, a move that would force index funds to automatically buy roughly $25B of the stock. The geographic map is shifting too. London's once-rough King's Cross neighborhood is now a global AI hub, hosting Google's UK headquarters alongside Anthropic, DeepMind, Synthesia, and Wayve. And Malta just became the first country to give every citizen a ChatGPT Plus subscription and an AI literacy course.

None of this surprises Charlie Stross, the science fiction author who wrote Accelerando back in 2005. He said this week that what he depicted in the book was simply extrapolation, because "the direction things were going in was obvious in the late 90s."

That's today. More tomorrow.