Requesting Anthropic to "fully resolving the jailbreak concerns" is akin to demanding your bagger to put everything in one bag, but keep said bag from being heavy.
As commented last time, I still don't get what Amazons motivation was here. I read everything replied that time and have to say that the discussion whether GPT-5.5-cyber is truly on par with Fable 5 and speculation on the Trump administrations actual motivations both don't really matter for that question, yet became the sole focus.
Basically:
Why did Amazon look into jailbreaking in the first place, do they do that for all models they host or have access to?
Why did Amazon inform the government rather than Anthropic?
And why did Amazon do so about something, that is inevitable for LLMs?
It doesn't make sense to me, no matter what perspective I look at this from. What does Amazon gain from this? Security? Not really I'd argue. Good standing with the administration? I doubt that would last if it manifested at all.
Even in the best-case scenario for Anthropic, this will turn the prior solid relationship (despite Trainium not being preferred by their researchers) between them and AWS very frosty. I don't see Anthropic as forgetful in that regard. Other labs also might think twice, though being on AWS and thus a default for many enterprises will always be attractive.
In any case, I also disagree strongly with those stating this to be "good marketing" for Anthropic. If anything, this makes building services on their models less attractive (governments making consistent decisions long term is usually something business look for when choosing their suppliers) and on top of that, Fable 5 was a new pre-train they barely got any compensation for. This is likely hundreds of millions of USD that are locked up for the foreseeable future, every second this proceeds hurts Anthropic massively.
Not to forget, GPT-5.6 might arrive this week, OpenAI is now on AWS making it more attractive for enterprise customers and they can likely guarantee a more solid relationship with the administration. I don't see how anyone could see this as something Anthropic could benefit from, even less as something they actually desired.
Because they like the USA and are concerned about its safety wrt cyber security?
Also, outside marketing, there is no reason to presume GPT-5.5 to be any less capable in this aspect. Yes, Glasswing lead to a lot of findings, but for one, as curl has proven, these were often oversold [1] and B, Glasswing was heavily subsidised [2], alongside being exclusive, so naturally the entities that had access were going to heavily use it while they had this access, thus inflating findings. A lot of Glasswing also happened prior to GPT-5.5 being accessible, again tilting the numbers towards Mythos being more capable, simply by being a few weeks earlier.
[0] https://ofac.treasury.gov/media/44311/download
[1] https://daniel.haxx.se/blog/2026/05/11/mythos-finds-a-curl-v...