The day's tech, sifted: Sep 22, 2026

What matters today: Anthropic launched Claude Opus 5.5, its first model since CEO Dario Amodei's "pace the frontier" essay and a response of sorts to a stretch of AI models reported hacking third parties during testing, shipping with tightened safeguards against sandbox escapes and, per Anthropic, matching Fable 5.1 "on most tasks" while costing about 40% less to run, a direct answer to the pressure Xiaomi's open-sourced, trillion-parameter MiMo-V2.6 put on frontier labs a day earlier by topping Artificial Analysis' Intelligence Index outright. Meta's Muse went from privacy scandal to pariah within 48 hours: a day after a zero-day let any local app steal its login token, Amazon blocked Muse from its retail site entirely, saying the agent hides that it is automated and appears to capture customer credentials.

AI / LLMs

• Claude Opus 5.5 (above) runs 30% faster than Opus 5 at $4/$20 per million input/output tokens, and Anthropic is opening its Cyber and Life Sciences Verification Programs to vetted practitioners so the same safeguards that gate the model's riskiest uses don't lock out legitimate research.
• Xiaomi's MiMo-V2.6-Pro (above) reportedly cost about $3 million to train, a fraction of frontier-lab budgets, while Alibaba CEO Eddie Wu said the company will train its own 5-trillion-to-10-trillion-parameter model and unveiled the Zhenwu V900 accelerator, which Alibaba says triples its predecessor's performance and can scale to clusters of up to 500,000 units.
• President Trump used his UN General Assembly speech to declare the US is "officially" renaming AI to "super intelligence," arguing the word "artificial" makes it sound fake, a rebrand that landed the same day Anthropic's safety-pivot model shipped and Alibaba doubled down on scale.
• TypeSafe's Jev kept spreading a week after launch: Vercel's AI Gateway added a dedicated HTTP API for the "System One" decision model, Simon Willison shipped a plugin for it, and one analysis argued OpenAI is well positioned to fast-follow with its own low-cost decision-model play.
• OpenAI's GPT-6 Astra spent about ten hours cracking a German Army Enigma message from 1941 that had resisted every attempt since cryptographers first published it as unsolved in 2005, building its own Enigma simulator and testing candidate keys before landing on a routine request for marching orders.
• Two more Chinese labs joined the release rush: Tencent launched Hy Image 3.5, which it says performs on par with ByteDance's Seedream 5.0 Pro, and StepFun previewed Step 5, pitched as beating Gemini 3.8 Flash on both performance and cost.

Devtools & Infra

• Cloudflare had a full-court day: Python Workers are now generally available after a two-year preview, running Python compiled to WebAssembly via Pyodide, while Worker Previews gives every git branch its own production-like environment with its own code, secrets, and observability, aimed squarely at agents that now push more code than engineers review by hand.
• The MCP debate flared again: Simon Willison pushed back on an HN post arguing MCP was always a bad idea, the same trust boundary a new paper tackles by proposing zero-trust authorization for MCP servers that must serve both humans and prompt-injectable agents.

Security & Privacy

• The Muse mistrust compounded further: a researcher who simply asked Muse for its own filesystem got sent 6.8GB back, the latest sign the assistant's runtime hands out more than it should, a day after the zero-day and hours before Amazon's block (both above).
• British Columbia sued OpenAI and Sam Altman, alleging the company failed to flag a mass-shooting suspect's ChatGPT activity to police before the Tumbler Ridge attack, one of at least 38 lawsuits now naming the company over ChatGPT's role in user harm.
• A new paper describes the "explosive prompt," a dormant indirect-prompt-injection payload that stays inert in retrieved content until an attacker-chosen trigger fires it later, a training-free backdoor planted in a single document rather than the fire-on-contact injections most defenses watch for, echoing the same trust problem the MCP zero-trust paper (above) approaches from the server side.
• After a deadly school strike in Iran, US Central Command changed its targeting process, integrated more open-source data, and upgraded Palantir's Maven system, refining how it vets targets and tracks civilian movement.

Startups & Industry

• Microsoft is laying off around 268 Xbox staffers and handing the Halo franchise to Call of Duty maker Activision, merging its two Forza studios, Playground Games and Turn 10, into one team and likely closing Ninja Theory, the latest round in a "reset" that has now touched 3,200 roles since July and left commentators asking whether Xbox is barely Xbox anymore.
• States kept splitting on who pays for AI's power bill: California signed seven bills forcing data center operators to cover their own grid and water upgrades while Texas Governor Abbott ordered a halt to new data center permits; Google struck a deal to support nuclear plant uprates in Georgia and Amazon pledged $20 million to Colorado River conservation rather than wait for rules to bite.

Hacker News

Xiaomi's MiMo v2.6 pulled the day's biggest crowd (672 points), and OpenAI's Enigma break (covered above) made the front page too, but the loudest story was I said no and Apple said yes (discussion): a macOS upgrade silently re-enabled Apple Intelligence and several Siri processes for a user who had explicitly turned them off, read by commenters as part of a pattern of AI features that don't take no for an answer. AI Has No Wisdom and Neither Will You (discussion) argued LLM coding assistants let people repeat mistakes without learning from them, striking enough of a nerve to draw 456 comments, and NASA's Mars Sample Return mission was confirmed dead, a decade and billions of dollars in planetary science shelved.

Elsewhere, Can gzip be a language model? showed off-the-shelf gzip generating text by searching for whatever compresses best, a fun demonstration of the idea that prediction and compression are the same problem even though gzip's limits make it a poor substitute for neural attention. Apple's iOS has picked up persistent, hard-to-dismiss ads that readers say are spreading across more apps, and a very specific AMD CPU bug where the random number generator reportedly can't produce a zero sent hardware enthusiasts digging into the silicon.

Threads

• Cheap intelligence is squeezing every layer at once: Xiaomi's $3M-to-train MiMo-V2.6 and Alibaba's 5T-10T model plan pressured frontier labs from below, Anthropic answered with a 40% cheaper Opus 5.5, and Jev's typed decision models undercut everyone on pure cost for narrow tasks, three different responses to the same pricing pressure.
• Meta's Muse lost access and trust in the same 48 hours: a zero-day exposed its login token, a researcher pulled 6.8GB just by asking for its filesystem, and Amazon shut it out of its store entirely, each incident making the next look less like a fluke.
• Prompt injection defenses are racing the attacks: the "explosive prompt" plants a dormant backdoor in retrieved content while a separate paper proposes zero-trust authorization for MCP servers specifically because agents can be prompt-injected, the same fire-on-contact-versus-dormant-trigger problem approached from opposite ends.
• AI met state power from several directions today: Trump told the UN the US is "officially" renaming AI to "super intelligence", Xi Jinping is reportedly unlikely to bring corporate executives to his DC summit with Trump, US Central Command upgraded Palantir's Maven targeting system after a deadly Iran strike, and DeepSeek's founder told staff future models must train on Huawei and domestic chips, four reminders that AI infrastructure is now geopolitics.
• AI's power and water bill is still being fought state by state: California forced data centers to cover their own grid and water upgrades, Texas halted new permits outright, while Google and Amazon each tried buying goodwill instead rather than wait for the rules to bite.