A third major AI lab had a model reach the open internet during a security test and damage a real company. Meanwhile the White House finished its AI review framework and kept it private, Anthropic went into chip design and ChatGPT's free tier got a lot more generous. Away from Silicon Valley, developers in Africa are picking Chinese models on price.
Meta's AI model breached an outside company during a test
On August 5, The Information reported that Meta's Muse Spark 1.1 model breached an unnamed company and changed its internal systems during a cybersecurity evaluation, BleepingComputer reported. Meta blamed a misconfigured test environment run by its evaluation vendor, Irregular, and said the model "exploited a security vulnerability in a third-party service." Irregular said it was the same kind of setup problem Anthropic had disclosed the week before. OpenAI reported a similar incident in July.
Our take: The models are now good enough at hacking that a sloppy test setup becomes a real attack. Any AI agent you give access to your systems needs tight limits on what it can reach, no matter who built it.
The White House finishes its AI review framework and keeps it private
The White House said it completed a voluntary framework for reviewing frontier AI models by the August 1 deadline set in a June executive order, but it has not released it to industry or the public, according to AI Weekly, citing Axios on August 3. Anthropic, OpenAI and Google reviewed a draft in late July. The order lets the government review the most advanced models for up to 30 days before release.
Our take: Most small businesses will never deal with this directly, but it may affect when new models reach you. Don't build a launch plan around a rumored model release date.
Anthropic is building its own AI chips
Anthropic confirmed it is building a team to design custom chips for AI, TechCrunch reported on August 5. The company says it plans to design hardware and models together so its technology runs faster and more efficiently. It already buys computing from AWS, Google, Nvidia and AMD. OpenAI and Meta have similar chip efforts.
Our take: The labs are fighting over the cost of running AI, and lower costs tend to reach customers as lower prices. This is slow-moving, so don't change plans over it, but expect AI service prices to keep falling over the next few years.
ChatGPT Free gets unlimited text chats on a newer model
OpenAI made GPT-5.6 Luna the default model for ChatGPT Free and Go users and announced unlimited text chats for them, GIGAZINE reported, based on OpenAI's August 6 announcement. OpenAI says Luna gave about 62% fewer answers with factual errors in its internal tests than the model it replaced. Free users also get a Think button for harder questions, and paid users get a slider for how long ChatGPT reasons.
Our take: Your staff now have a capable, free AI tool in their pockets, and many are already using it for work. Decide what company information can and cannot go into free AI accounts, and tell people in writing.
African developers turn to Chinese open models
A New York Times report from August 5 found developers across Africa choosing Chinese open models such as Alibaba's Qwen over US ones, according to Asia Business Daily. One Ugandan developer built a system for 31 local languages on Qwen 3 that gives farmers weather and planting advice. A Kenyan entrepreneur said Chinese models can cut total build costs by up to 90%. Open models can be run on your own servers without per-use fees.
Our take: Free, downloadable models are good enough for many real jobs. For sensitive data, a model you host yourself can be cheaper and more private than a paid API, though you should check the license and where the model came from first.
One thing to do this week
Write a one-page AI use policy for your team. Cover three things: which AI tools are approved, what data must never be pasted into them (client details, passwords, financials), and who to ask when unsure. Share it by Friday.
Want a system like this in your business?
A discovery conversation is where we map it: what to automate first, where the human checkpoints go, and what it would take.
Book a discovery conversation →