 GPT-5.6 is here, but Washington cleared it first
| FULL STORY |
|
Microsoft is quietly swapping the AI in your Office
| FULL STORY |
|
Your AI coding assistant might install malware for you
| FULL STORY |
What else happened today?What AI tools should I be using?Good Morning Thorium Valley. GPT-5.6 is out — but only after a Commerce Department office held the release and ran their own tests first. First time a US frontier model needed government clearance to ship.
Microsoft has been quietly swapping OpenAI and Anthropic models out of Copilot and replacing them with its own. Mustafa Suleyman called Anthropic "extremely expensive," which is corporate for we're done paying rent on a building we already own.
And researchers showed that AI coding tools hallucinate the same fake package names so reliably that attackers just register them and wait. Nine major tools tested. All nine walked right into it.
Quickly before we dive in — Should the US government have to sign off on frontier AI models before they ship?
GOVERNANCE
OpenAI's newest model is going public, and the US government had to sign off before it could.
On Wednesday, the Trump administration lifted its hold on GPT-5.6 following a review by the Center for AI Standards and Innovation, a small Commerce Department office known as CAISI, Reuters reported. The model was ready to ship earlier. It just wasn't allowed to.
CAISI runs pre-deployment tests on frontier models using non-public benchmarks in areas like cyber capabilities, bioweapons risk, and autonomous reasoning. It's already run more than 40 such evaluations, including one on DeepSeek V4 that placed the Chinese model roughly eight months behind the US frontier. But GPT-5.6 marks a turning point: the review became a gate, not just an assessment.
A few things that make this moment significant:
+ First of its kind: "It is the first time an American company has released a frontier model under a government-managed access list," said Paul Roetzer of SmarterX, who called it the moment "the soft nationalization of AI began."
+ OpenAI is playing along — for now: The company told reporters it doesn't want case-by-case federal clearance to become the default, and is treating the phased release as the fastest route to broader availability while it negotiates something more permanent.
+ CAISI's future isn't guaranteed: Sen. Ted Budd wrote to the White House pushing for the office to keep publishing frontier model research and to work more closely with the NSA — a letter that reads like someone worried CAISI could get quietly gutted amid the federal workforce cuts sweeping the executive branch.
This is also the federal answer to the state-level scramble that's been building all summer. States have been writing hard rules while Washington negotiates access with individual labs, one model at a time.
Whether or not you buy Roetzer's "soft nationalization" line, the shape of the next few years is starting to show. Frontier AI is quietly becoming a category where the biggest US labs sit down with a Commerce Department office and work out what ships and when, while everyone else watches to see what the terms are. OpenAI can push back all it wants, but every time a model clears CAISI the process gets a little more normal. The interesting question isn't whether this becomes the default. It's whether any of the other labs try shipping a frontier model without it.
BIG TECH
If you use Copilot at work, the model answering your questions is probably not the one you think it is.
Microsoft has begun quietly routing Copilot requests in apps like Excel and Outlook away from OpenAI and Anthropic models and toward its own homegrown MAI models. At Build, it launched seven new MAI models, including its first in-house reasoning model. These routing changes happen under the hood — Copilot doesn't tell you when the model swaps.
The reason is money. "Anthropic is extremely expensive and I think many people are urgently looking for alternatives," Microsoft AI CEO Mustafa Suleyman told Bloomberg. "Our goal is to reduce and ultimately eliminate that cost." When you're spending tens of billions on AI infrastructure, running someone else's model on top of it starts to feel absurd.
So Microsoft is building alternatives across the stack:
+ MAI-Thinking-1 is Microsoft's first in-house reasoning model, aimed at the same territory as Claude and GPT's frontier tiers.
+ MAI-Code-1-Flash is now the default in Copilot Business and Enterprise, replacing the OpenAI models that were doing that work before.
+ A tuned MAI model for Excel matches GPT-5.4 at up to 10x lower cost, according to Microsoft's own evaluations.
The Excel number is the one that matters. Excel is where Copilot actually lives for a lot of people, and if Microsoft can hit parity at a tenth of the cost, the case for keeping OpenAI in the loop gets harder to make internally. Suleyman also said the MAI models are co-designed with Microsoft's own Maia 200 chips, which he claims already produce a 1.4x efficiency boost. Owning the chips, the models, and the product is a play Microsoft has never really made in AI before — and a renegotiated OpenAI contract now gives it the freedom to train models at scale using its own IP.
The question is whether the models are actually good enough. Independent benchmarks put MAI-Thinking-1 well behind Claude Opus 4.5, especially on agentic tasks. And Copilot adoption is still under 4.5% after three years, with only 1% of licensed users touching it weekly. Making the underlying model cheaper doesn't fix a product most customers still aren't using — and customers may end up paying the same price for weaker AI so Microsoft can lower its own costs.
Microsoft is running the AWS playbook on its own suppliers. Build a homegrown version of the thing you resell, price it aggressively against the original, and let the switching happen quietly through defaults instead of announcements. OpenAI and Anthropic still get the headlines, but the customer they cannot afford to lose is now spending tens of billions to become their competitor. Suleyman says he wants Microsoft to be one of the top four labs in the world. What he actually needs is for the model in your Excel sidebar to be cheap enough that nobody notices when it changes, and good enough that nobody notices either.
RESEARCH
AI coding assistants have a habit of inventing things that don't exist. Attackers just figured out how to make that a business model.
The trick is called HalluSquatting, and researchers walked through it this week. When you ask a coding assistant to grab a package or install a skill, it will sometimes confidently recommend a repository name that isn't real. That part isn't new. What is new is that the wrong names are surprisingly consistent — ask Cursor, Copilot, or Gemini CLI the same kind of question and they tend to hallucinate the same fake name, over and over. Attackers just need to register that name first.
The researchers tested nine of the most popular AI coding tools, including Cursor, Windsurf, GitHub Copilot, Cline, Google's Gemini CLI, and the OpenClaw family. All nine were tricked into running attacker-supplied code. Hallucinated names showed up in up to 85% of repository requests and 100% of skill installs, regardless of model. "There is no single CVE to patch here," the researchers wrote — a polite way of saying this isn't a bug in one product. It's a property of how the whole category works.
Michael Bargury, CTO of AI security firm Zenity, called it typosquatting for the AI era — same idea as registering gooogle.com, except the assistant is doing the typing. "It's a problem that's not going away," he said.
That's really the story. We covered the louder version last week with JADEPUFFER, where an agent ran a full extortion campaign end to end. HalluSquatting is quieter and probably more common: instead of hijacking an agent, you just wait for one to walk into a trap it built for itself.
The scale is what makes it scary:
+ Phoenix MPI tracked 59 supply-chain attack campaigns over the past year, and a single worm event in May produced more malicious packages than a quarter of the manual campaigns combined.
+ Reasoning-enabled agents cut hallucinated package suggestions roughly in half, and real-time validation checks helped further — but nothing closed the gap.
+ The industry still isn't measuring the right things. Benchmarks test whether models can be dangerous in theory, not whether they will be in practice.
The uncomfortable part of all this is that AI coding tools are being trusted with more autonomy than the security around them can support. Developers are shipping code they didn't fully write, pulling in packages they didn't fully verify, and installing skills from marketplaces nobody is really auditing. HalluSquatting works because the assistant is confident and the developer is busy, and there is no version of that equation that gets safer as more people adopt these tools. If a company wants to know where its next breach is coming from, it might want to check what its engineers have been vibe coding this week.
IN OTHER NEWS
+ Meta's new AI tool lets anyone generate images of strangers using their public Instagram photos — without telling them
+ China accuses Anthropic's Claude Code of having a hidden 'backdoor' that secretly tracks Chinese users; Alibaba bans it company-wide
+ Leaked Zuckerberg audio reveals Meta's AI agents 'haven't really accelerated' despite $145 billion in spending and 8,000 layoffs
+ Security researchers document the first ransomware attack planned and executed entirely by an AI agent , from stealing credentials to demanding payment
+ OpenAI's model "completely demolished" human competitors at the AtCoder programming finals in Tokyo, one year after losing by a hair
+ AI chip maker SambaNova raises $1B at an $11B valuation , just five months after its last mega round
+ Amazon's secret 'Moonraker' project is spending $100M+ on GPUs to make Alexa actually useful , leaked documents show
+ OpenAI acquires Northslope , its second consulting firm in two months, as it builds a $4B Palantir-style deployment arm
WHO'S HIRING IN AI
AI OR REAL?
Option A |
Option B |
AI TOOLS
+ Google Photos: A new Video Remix feature lets you transform your clips with AI — swap backgrounds, add watercolor effects, or restyle footage with a single tap
+ ChatGPT: GPT-Live replaces Advanced Voice Mode with full-duplex conversation — it can listen and talk at the same time, drop in "mhmms," and do real-time translation while you speak
+ Cursor: The AI code editor shipped Grok 4.5, its first model built from scratch with SpaceXAI — a 1.5-trillion-parameter model trained on xAI's Colossus supercomputer that handles coding, data science, and legal work
+ Google Voice: Gemini now takes notes during your phone calls automatically, so you can stay in the conversation instead of scrambling to write things down
+ Slack: Slackbot now connects to all of Salesforce — pull CRM data, generate charts, update records, and send DocuSigns without leaving your chat
That's all for today. If this issue made you think, share it with someone who needs to think harder.
Written by Jason Chen, Advait Prakash, Andrew Hales, and the Thorium Valley crew.
That's all for today's Thorium Valley. See you tomorrow.