Get tomorrow's brief in your inbox
Today: Anthropic's Fable 5 and Mythos 5 models were banned by the U.S. government following security concerns raised by Amazon, shutting off access worldwide. KPMG pulled an AI-usage report filled with hallucinations, and OpenAI is now under investigation by state attorneys general over safety and consumer protection issues.
The U.S. government ordered Anthropic to completely cut off access to its Fable 5 and Mythos 5 models on Friday, blocking all foreign nationals from using them, including Anthropic's own employees. The directive came after Amazon CEO Andy Jassy reportedly briefed Treasury officials on security research showing the models could be jailbroken to provide information useful in cyberattacks.
Anthropic disputes the severity, arguing the vulnerabilities are minor and exist in other public models like GPT 5.5. The company said it received only verbal descriptions of concerns, not detailed disclosures. This is the latest escalation in a running battle between Anthropic and the Trump administration over the company's refusal to allow AI for mass surveillance or lethal autonomous weapons. (TechCrunch, The Verge)
If you're a U.S. business currently using Claude Fable or Mythos, you're affected. The models are offline globally. Anthropic's older models (Opus, Sonnet, Haiku) remain available, but this incident highlights real risks when relying on a single AI provider. The takeaway for SMBs: diversify your AI stack. Have backup options configured, whether that's OpenAI, Google's Gemini, or open-source models you can run locally. Vendor lock-in is now a national security risk, not just a business continuity problem.
India's tech sector is rethinking its dependence on U.S. AI providers after the ban also hit companies there. Anthropic had just partnered with Tata Consultancy Services to expand in India, calling it their second-largest market after the U.S. Indian founders are now pushing to accelerate domestic AI development and open-source alternatives. (TechCrunch)
Python developers can now publish WebAssembly (WASM) packages directly to PyPI for use in browser-based Python environments like Pyodide. This removes a longstanding distribution bottleneck where Pyodide maintainers had to manually build and host over 300 packages themselves. Package authors can now compile C or Rust extensions to WASM and distribute them just like native wheels. (Simon Willison)
If you're building web-based tools or internal dashboards that need Python in the browser, this matters. You can now use more Python libraries without waiting for Pyodide maintainers to package them. Simon Willison demonstrated by publishing luau-wasm, a Lua-based scripting engine, to PyPI in under a day. There are currently only 28 packages using this format, but expect rapid growth as the word spreads.
Apple shipped three AI photo editing features in the iOS 27 developer beta: Clean Up (remove objects and people), Extend (expand photo edges with AI filler), and Spatial Reframing (recompose shots by simulating camera movement). Clean Up finally works well after being upgraded to use cloud models instead of only on-device processing. (The Verge)
For marketing and social media teams, this puts Google Pixel-level editing tools in the hands of the majority of smartphone users. The quality is good, but the bigger story is what this does to trust in photography. Apple's Spatial Reframing can change your perspective on a scene you already photographed. That's powerful for fixing composition mistakes, but it also means photo evidence is becoming less reliable. If you run a business that depends on photo authenticity (real estate, insurance, legal), you need verification workflows that go beyond "send me a picture."
A non-technical user built a custom yard work app using Google's Gemini AI by writing a single lengthy prompt describing what they needed: manage chores, recommend tasks based on weather, and diagnose plant problems using image recognition. Gemini generated a working Android app preview in minutes, complete with live weather API integration and an AI plant doctor feature. (The Verge)
The user spent more time arguing with Gemini about color schemes and UI details than writing code. This is the real promise of AI coding tools: turning subject matter experts into app builders without requiring them to learn JavaScript. If you have a niche workflow problem that off-the-shelf software doesn't solve, try describing it to Claude, ChatGPT Code Interpreter, or Gemini's app builder. You might get something useful in an afternoon that would have taken a developer a week to spec and build.
The lesson isn't that AI writes perfect code. It doesn't. But it writes working code fast enough that iterating on bugs and design is faster than waiting for a dev shop quote.
Startups are using AI video tools to create marketing content faster and cheaper, fundamentally changing how they test messaging and audience fit. The shift isn't about production quality. It's about iteration speed. You can test 10 different video pitches in the time it used to take to produce one polished piece. (Inc.com)
If you're still scripting, shooting, and editing one perfect video at a time, you're doing it wrong. Use AI tools (Runway, Synthesia, HeyGen, or even ChatGPT-generated scripts with stock footage) to create 5-10 variations and let your audience tell you which one works. The best-performing version can then get the professional treatment. This is the lean startup method applied to content: test first, polish later.
Meta is dismantling its $2 billion acquisition of Manus, the AI startup formerly known for viral agent demos, after Beijing issued a divestiture order on national security grounds. The company has cut Manus off from internal systems and stopped all data sharing. Manus co-founders are now in talks to raise $1 billion to buy the company back from Meta and restructure it for a Hong Kong listing. (TechCrunch)
This is part of a broader Chinese government crackdown on AI exports. Top AI firms including Moonshot AI and StepFun now need government approval before accepting U.S. investment. Travel restrictions have also been expanded, requiring researchers and executives at private AI firms to get approval before going abroad.
For businesses, the message is clear: global AI supply chains are fragmenting along geopolitical lines. If you're building products that depend on Chinese AI companies or serving customers in China with U.S.-based AI, expect friction. Plan for regulatory delays and possible service interruptions.
A coalition of state attorneys general has opened an investigation into OpenAI, with New York's AG serving a subpoena covering advertising practices, user retention tactics, model sycophancy, data handling, and treatment of minors and seniors. The move comes as OpenAI faces multiple lawsuits over copyright infringement, user safety, and alleged failures to alert law enforcement when flagging dangerous account activity. (TechCrunch)
OpenAI recently announced it filed confidentially to go public, which makes this investigation particularly awkward timing. For businesses using ChatGPT or OpenAI APIs, this doesn't change anything operationally in the short term, but it's another signal that regulatory scrutiny on AI companies is intensifying. Expect more transparency requirements, age verification mandates, and potential restrictions on how AI tools can engage users, especially younger ones.
Professional services firm KPMG retracted a report on AI usage after organizations named in the report said the claims were false. The report, titled "Redefining excellence in the age of agentic AI," appears to have used AI to write content about AI, resulting in fabricated case studies. UBS, the UK's NHS, Swiss Federal Railways, and Transport for London all told the Financial Times that KPMG's claims about their AI deployments were untrue or misleading. (TechCrunch)
This follows EY pulling a similar report on loyalty programs last month due to fake footnotes and hallucinations. The pattern is concerning: if Big Four consulting firms can't validate AI-generated reports before publishing them, what does that say about due diligence standards elsewhere?
If you're paying consultants or agencies to deliver AI-assisted work, add verification steps to your contracts. Require that any AI-generated claims be backed by primary sources you can verify independently. The "AI wrote it and we shipped it" approach is becoming a liability issue.
Diversify Your AI Provider Stack Today
Don't wait for your primary AI tool to get banned or go offline. Set up accounts with at least two alternative providers right now. If you use ChatGPT, configure Claude or Gemini as a backup. If you use Claude, have OpenAI API keys ready. Test your critical workflows on both systems so you know switching costs before you're forced to switch in a crisis.
Audit Your AI-Generated Content for Hallucinations
If you've published reports, case studies, or marketing materials with AI assistance, spot-check the factual claims. Call the companies you mentioned and ask if the case studies are accurate. Check that statistics have real sources. KPMG and EY learned this lesson the expensive way. You don't have to.
Geopolitical risk is now an AI vendor risk. This week's Anthropic ban and Meta's forced Manus divestiture show that access to frontier AI models can disappear overnight for reasons that have nothing to do with your business. The U.S. and China are fragmenting the global AI market, and companies caught in the middle are paying the price.
Meanwhile, the accountability gap is closing. KPMG's hallucinated report and OpenAI's attorney general investigation signal that "the AI made a mistake" is no longer an acceptable defense. Regulators and customers expect human oversight and verification. If your business uses AI for anything customer-facing or compliance-critical, build verification workflows now before you're the next case study in an investigation.
Watch for fallout from the Anthropic ban in India's tech sector. If major markets start prioritizing domestic or open-source AI over U.S. models, that reshapes the competitive landscape for everyone.