Get tomorrow's brief in your inbox
Today: The US government shut down public access to Anthropic's most powerful AI models over alleged security concerns, while Google sued a Chinese cybercrime network using AI to scam hundreds of thousands of victims. OpenAI launched free training courses to help companies build AI fluency across their workforce.
The US government ordered Anthropic to immediately disable its two most powerful AI models, Claude Fable 5 and Claude Mythos 5, citing national security concerns over a claimed jailbreak that could identify software vulnerabilities. The directive, issued Friday at 5:21 PM ET, forces Anthropic to cut access for all users worldwide, not just foreign nationals.
Here's what's bizarre about this: Anthropic says the government's concern is based on a technique that amounts to asking the model to read code and find bugs. That's exactly what cybersecurity professionals do every day, and other publicly available models like OpenAI's GPT-5.5 can already do the same thing. Anthropic argues that if this standard were applied across the industry, it would "essentially halt all new model deployments for all frontier model providers."
For small businesses, this is a warning shot about AI regulation. If you're building business processes around cutting-edge AI models, you need fallback options. Anthropic's other models remain available, but this shows how quickly government action can disrupt your tools. The practical takeaway: don't build mission-critical workflows around a single AI model. Have alternatives ready.
Source: TechCrunch | Source: Simon Willison
OpenAI Academy expands with three free courses
OpenAI launched three new training courses designed to help your team actually use AI in daily work: AI Foundations (basic prompting and responsible use), Applied AI Foundations (building repeatable workflows), and Agents and Workflows (directing agent-assisted work). You complete hands-on exercises using your own work, not generic examples. Learners get certificates they can share, which helps you track who's trained across your organization.
This is free training built by the company that makes ChatGPT, and it focuses on practical application rather than theory. If you're struggling to get your team beyond "I asked ChatGPT to write an email," these courses provide a structured path from basic prompting to building workflows that save real time. Sign up at OpenAI Academy.
Google rolls out AI Inbox and productivity features
Google announced AI Inbox for Google AI Ultra subscribers (rolling out to Plus and Pro users soon). It highlights urgent emails, generates personalized draft replies, and integrates relevant Google Docs, Sheets, or Slides directly into your task list. You can mark tasks done, dismiss suggestions, or bulk mark related emails as read. Gmail Live adds voice commands for email management.
Also new: Google Pics creates and edits images without design expertise (limited testing, global rollout this summer). Docs Live lets you create documents through voice commands, summarizing and organizing as you speak. Keep now has voice-to-text for transforming brainstorming into organized notes.
If you're already in the Google ecosystem, these tools are coming to you. The AI Inbox feature is the most immediately useful for solopreneurs drowning in email. Worth trying when it hits your account.
OpenAI upgrades WebRTC audio with document context
OpenAI updated its WebRTC audio API to include GPT-Realtime-2, their first voice model with GPT-5-class reasoning. You can now paste document context and have an audio conversation in your browser about that information. The tool is free to use at Simon Willison's playground.
This is useful if you learn better by talking through concepts than reading. Load a technical document, spec sheet, or report and have a conversation about it. Not a replacement for reading, but a different way to process complex information.
AI governance for small businesses
Zapier published a guide to the six best AI governance tools: Zapier (for building safely with AI), Microsoft Purview (for Microsoft users), OneTrust (enterprise-level governance), Credo AI (regulation management), Fiddler (real-time observability), and ModelOp (for regulated industries).
AI governance covers three areas: monitoring how employees use AI and controlling data access to external models, complying with regulatory frameworks like the EU AI Act, and preventing issues like prompt injection and data leakage in production. For most small businesses, the first category matters most. If your team is using ChatGPT, Claude, or other AI tools, you need visibility into what data they're sharing.
Start with Zapier's AI governance features if you're already using their platform. For Microsoft shops, Purview is built in. If you're in a regulated industry (healthcare, finance), read this guide carefully before deploying AI tools.
Jeff Bezos building "artificial general engineer"
Jeff Bezos announced his new AI startup Prometheus raised $12 billion at a $41 billion valuation. The company aims to build AI-powered engineering tools for robotics, drug design, and manufacturing. Bezos serves as co-CEO alongside Vik Bajaj from Alphabet's Verily. The company has 150 employees.
Bezos says Blue Origin (his rocket company) is a perfect example of a company that could benefit: "Any company that is building sophisticated devices like rocket engines would benefit greatly from this kind of technology." The tools are aimed at physical product development, not software.
If you're in manufacturing, this is a market signal that AI tools for physical product design are coming. Don't expect them this year, but start thinking about what you'd do with an AI that could help design parts, optimize layouts, or suggest manufacturing improvements.
Google Calendar appointment scheduling vs. Calendly
Google Calendar's free appointment scheduling feature has improved enough to compete with Calendly's free tier. Both limit you to one booking page and one calendar connection on free plans. Calendly's edge: it auto-generates links for Zoom and Teams, not just Google Meet, and has a mobile app.
Calendly's paid plans start at $12/user/month (unlimited event types, multiple calendars, payments, automated reminders) and $20/user/month for teams (enterprise integrations, round-robin meetings, lead routing). Google Calendar's appointment scheduling is free with any Google account, but premium features require Business Standard plans or higher.
If you're a solopreneur with simple scheduling needs and you already use Google Calendar, try the built-in appointment scheduling before paying for Calendly. If you need payment collection, custom branding, or multiple booking types, Calendly's Standard plan at $12/month is worth it.
Google sues Chinese AI-powered cybercrime network
Google filed a lawsuit to dismantle Outsider Enterprise, an alleged Chinese cybercrime operation that used AI to scam "hundreds of thousands of victims" with losses estimated in the millions. The group deployed 9,000 fake websites, one million fraudulent domains, and sent 2.5 million texts to Android users in two weeks.
The operation sells phishing software called Outsider for $88/week or $200/month. It creates fake websites with AI (including Google's own Gemini) impersonating telecom providers, banks, government agencies, and retailers. The platform offers 290 pre-built templates and guides on how to "weaponize AI-generated code." Cybercriminals coordinate on Telegram channels.
Google says it intercepts over 10 billion scam messages a month using AI-powered detection. The FBI confirmed it seized several domains and Shopify storefronts used by the operation, with an estimated 3.87 million stolen credit cards and $1.9 billion in losses since July 2023.
For your business: if you're using AI to generate code or automate processes, assume bad actors are doing the same at scale. The barrier to launching sophisticated phishing campaigns is now $88 and basic technical skills. Train your team to verify before clicking.
Meta's Applied AI team in turmoil
Meta's three-month-old Applied AI team of 6,500 engineers is reportedly on the verge of revolt. Employees were drafted into the unit through surprise emails with a choice: join or quit. Their work: generating puzzles and coding problems to train AI models. Employees call it "soul-crushing" and describe themselves as "draftees."
CEO Mark Zuckerberg's reasoning: Meta employees have "significantly higher intelligence" than third-party contractors for data labeling work. Over 1,600 employees company-wide signed a petition protesting a program that monitors their clicks and keystrokes for AI training data.
If you're wondering why AI companies are struggling to hire, this is a data point. The grunt work of training AI models is tedious, and top engineers don't want to do it. For small businesses, this means AI tools will continue improving, but maybe slower than the hype suggests.
Talkspace launches AI therapist with major limitation
Talkspace launched an AI therapist, but the article content is sparse. As AI chatbots face lawsuits over offering clinical advice, this is a notable move by a telehealth company. Details on pricing, safeguards, and use cases weren't available in the source material.
Approach this category carefully. AI for mental health support is a minefield of liability and ethical concerns. If you're in healthcare or wellness, wait for clearer regulatory guidance before deploying AI therapist tools.
Oracle PeopleSoft critical vulnerability actively exploited
If you use Oracle PeopleSoft, patch immediately. The ransomware group ShinyHunters exploited a critical vulnerability (CVE-2026-35273, severity 9.8/10) for over two weeks before Oracle flagged it. About 100 organizations were targeted, with 68% in higher education. The University of Nottingham confirmed significant student data was stolen.
Oracle issued a stopgap mitigation but hasn't fully patched the flaw yet. If you run PeopleSoft, apply the mitigation today and monitor for Oracle's full patch.
The government's shutdown of Anthropic's most capable models signals that AI regulation is moving faster than the industry expected. At the same time, AI-powered cybercrime is industrializing with turnkey phishing platforms available for under $100/week. For small businesses, the pattern is clear: AI tools will continue getting more powerful, but access could be restricted with little warning, and the criminal use of AI is accelerating. Focus on building processes that can adapt when tools change, and invest in security awareness training for your team. The tools are getting better, but so are the threats.