what is actually going on
To start out this blog post I want to give some background around the entire situation and why things are how they are right now. It all basically started April 7 2026 when Anthropic released a new project called Glasswing this project gave big corporations like AWS, Apple, Broadcom, and a few others access to the new Mythos Preview Model, the important part thought is the way Anthropic framed the entire situation, the tldr is that Mythos preview was way too dangerous to release since it is too good at cyber security showing charts like this:
or quoting 3 big vulnerabilities they found. This is in my opinion meant to make people fear the model and since they at that time only released it in Glasswing not to the general public. Exactly 63 days later Anthropic released the two models Mythos 5 and Fable 5, these are the same model just with different safeguards applied to it. 72 hours later the model was taken offline to the general public as the US government issued an export control directive to suspend access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the US. About 18 days later access to these models was restored with heavier safeguards in place. These safeguards are meant to trigger more on benign requests to make sure fewer harmful requests manage to go through as seen here:
I hope this gives you some insight into the situation before I give my take on it and some more things I personally think.
anthropic got this on themselves but doesn't want to cooperate
Some insight into why the model got export controlled is that Amazon employees[1] used the Model Fable 5 and asked it to find bugs in a codebase, since the model isn't meant to actually do cyber security things as that would allow abusing it, it shouldn't do this. The model complied and found bugs in the codebase, Amazon Employees brought this up to Anthropic who did not do anything in this situation, now you have to decide yourself if you would consider this a jailbreak but in my opinion this is just expected and wanted behaviour because otherwise the models would be quite useless. Following this it got escalated all the way up the chain to the Amazon CEO which then himself called up the treasury director, this then directly caused the export directive to be issued. This is not the first incident of Anthropic with the US government where in March they got classified as a supply chain risk to America's national security after refusing to supply the Department of War with their AI for the use of autonomous weapons. [2]
I personally think this is ironic in the way that all the fearmongering that Anthropic has been doing harmed themselves.
anthropic is interfering even with other labs
On June 26 2026 OpenAI released the models GPT 5.6 Luna, Terra and Sol. These models had a similar launch to Mythos by anthropic being available to only trusted partners.
During the preview, GPT‑5.6 models will initially be available through the API and Codex to a select group of trusted partners and organizations. We plan to make them more broadly available to people using ChatGPT, Codex, and the API soon.[3]
At the date of writing this post the models are not yet publicly available. Anthropic has publicly stated that they want a slowdown of all AI labs to deal with implications. They stated that if this just let less cautious actors catch up it would leave everyone less safe. Voting towards a global coordination mechanism. Anthropic has done this kind of fearmongering for the last while in more cases than I just cited.
it is just fearmongering the models aren't that harmful
A big point in this is that even though Mythos is a good model Anthropic is exaggerating it such as seen here, this is a post about a Mythos scan that apparently found 5 curl vulnerabilities, of which in the end only one was a low-severity CVE. This is in my opinion quite a big point on how scary things will look if you don't look deep enough.
what this could do to the ai landscape
As we see here the things anthropic has been doing have definitely interfered and changed some things in the AI landscape, I personally doubt we will get a full model release by any major lab in the next time, I think most are gonna be these staggered releases to trusted partners first. The issue with this is that Chinese labs such as Z.ai with very powerful openweight models such as GLM-5.2 will most likely not have these kind of releases. I think if Anthropic will make the major labs put safeguards in front of models and do these releases this could cause people to just stop using them and lean towards openweight which means that those models don't have safeguards and in the end this could cause the entire AI landscape to become less safe than it was all along. This is also the case since we already have releases with 2 seperate models like with Fable 5 and Mythos 5 where the actual publicly released model is heavily guarded and has forced 30 day data retention. Which makes the model inaccessible to big companies since those often require 0-data-retention by policy. Right now the issue is that we don't really have a choice, Fable/Mythos 5 is a really good model, it is the best model[4] (This post was written before GPT 5.6 was publicly released). Yes GLM-5.2 is an awesome model but compared to both Fable 5 and GPT 5.5 it is not yet good enough to be able to just use openweight models. We rely on US ai labs and if anthropic ruins it this is a huge step in the wrong direction.
conclusion
In conclusion I just want to say that this has definitely gotten out of hand and I hope either openweight models manage to actually catch up or Anthropic finds another way to keep AI open, as was the original goal with the company OpenAI. I believe it is better to have safe good models with safeguards that don't interfere with everyday work than worse openweight models that have none.
Footnotes
-
Wall Street Journal — Amazon CEO's Talks With U.S. Officials Triggered Crackdown on Anthropic Models ↩
-
The Blog Post from Anthropic about this Situation - Where things stand with the Department of War ↩
-
DeepSWE — a long-horizon software engineering benchmark. As of July 1 2026, claude-fable-5 sits at the top of its leaderboard with a 70% pass@1 rate, ahead of gpt-5.5 at 67%. ↩