Hello, dear readers!

This is our weekly brief on remarkable AI topics, so you can keep up without drowning in the noise.

Today’s focus — Anthropic's Fable is back, now with maybe too many safeguards. The company's most advanced models briefly wandered into export-control territory, then returned with a lot of caution tape and a more nervous door policy. The question is no longer just whether a model is smart, but who is allowed to use it, where, and under what conditions.

Also in this week’s edition:

Anthropic escapes model jail

Anthropic’s most advanced models are finally back — but not quite as they were. After a brief period of restricted access over cybersecurity concerns, Fable 5 and Mythos 5 have returned with tighter controls, added safeguards, and a lingering sense that they are now being watched more closely.

The underlying issue, according to Anthropic, was less about unique dangerous capability and more about perception. In its tests, other models could identify the same vulnerabilities and even produce similar exploit examples. What made this different was visibility: a high-profile model triggering a high-level response.

To bring the models back, Anthropic added a new safety classifier designed to block the technique described in the Amazon report in more than 99% of cases. If a request gets flagged, it is blocked or routed away from Fable. The company also acknowledges the tradeoff: the system may catch some benign coding and debugging requests along the way.

So Fable escaped model jail, but it did not return as a carefree machine god. It returned with a stricter bouncer — and some users noticed. Early reactions have been mixed: the model is still powerful, but there are complaints that it can feel more cautious, more constrained, or less immediately useful than the forbidden-object drama might have suggested.

When the models were re-released, some users may have expected forbidden intelligence. What they got was something closer to a very powerful assistant standing behind several layers of caution tape.

AI used to ship like software.

Now it increasingly ships like infrastructure with a national security appendix.

China drops cheap AI dragon

While the U.S. was busy deciding who should be allowed to touch Anthropic’s models, China produced a different kind of headache for the AI race.

Z.ai, the company formerly known as Zhipu AI, has released GLM-5.2, a Chinese open-weight model that is getting attention for strong coding and agentic performance at a much lower cost than leading U.S. models.

The pitch is brutally simple: what if “frontier-ish” intelligence did not require frontier pricing?

That matters because AI competition is no longer just about who has the best benchmark chart. It is also about distribution. If a model is good enough, cheap enough, and easy enough to deploy, developers will try it. Startups will try it. Companies priced out of premium U.S. models will definitely try it.

Of course, the “Chinese open-weight model” part comes with its own baggage. Western companies, especially in regulated industries, may worry about data security, geopolitics, compliance, and where exactly the model is being run. But the fact that those questions now have to be asked is the point.

The AI race is becoming less tidy. The frontier is no longer only a handful of closed American labs with expensive APIs. There are open weights, cheaper challengers, national champions, local deployments, cloud wrappers, and developers who mainly want something that works without destroying their budget.

For OpenAI and Anthropic, that is annoying.

For everyone else, it may be the most interesting part of the week.

Zuck discovers agents are hard

And then there is Meta, where the AI future has run aground on the oldest obstacle of all: implementation.

According to Reuters, Mark Zuckerberg told employees that the company’s AI agent development has been going slower than expected.

Meta has spent the past year talking about agents as one of the next major steps in AI — software systems that do not just answer questions, but perform tasks on behalf of users.

The reality is harder. Agents need memory, tools, permissions, reliability, user trust, error recovery, and some way of not turning every multi-step task into a workplace disaster. A chatbot can be charming while being wrong. An agent being wrong is more like giving a confused intern access to your inbox, credit card, and production database.

That does not mean agents are dead. It probably means they are entering the boring, expensive, and necessary phase of becoming real products. Useful systems move slower than demos because they have to survive contact with users, edge cases, and company policies.

Until next Monday — by then, an AI agent will almost certainly have scheduled three meetings, cancelled two of them, and confidently joined the wrong Zoom.