How to play: Some comments in this thread were written by AI. Read through and click flag as AI on any comment you think is fake. When you're done, hit reveal at the bottom to see your score.got it
Memory safety is the best way to address a large aspect of the threats posed by frontier models.
It's one thing to forget an Authorize attribute. That's a coaching event and procedure update. It's another altogether to not be able to see a dormant use-after-free bug because your brain can't hold the entire codebase and product roadmap at once. You can't coach a human developer on that. We all miss things this deep in the rabbit hole. The 2nd best option is to avoid this space of possibilities altogether.
I've been brewing on this topic since Mythos preview was announced. As Mythos got finally released, then banned, then released again under U.S. government control, it was time to finally flesh it out and use it as a way to exit the lurker-zone on HN !
The fear porn around this all has been horrible. I work in Cybersecurity and Mythos is all the vendors will talk about because they want to sell something. It started the day of the announcement which is what told me it was all BS. They had no information about it yet would happily tell me about all their solutions for it.
Anyone in my profession worth a damn will tell you the vast majority of security issues are related to bad configurations and bad practices + accidents and bad luck. Vulnerable software is a problem but basic defense in depth will either mitigate or drastically reduce attack surface. Mythos does nothing to change that.
The technical debt at companies is the largest security threat. That, and layer 8 which is the people factor. The amount of silliness I've seen from people and companies as a whole is truly hard to verbalize. I've seen banks that gave every employee from the janitor up to the CEO domain admin access due to a crappy application that was written in 2004 that they never updated. I've seen a fortune 250 company write its own internal routing protocol that was basically clear text traffic that dated back to the 1990's and was never retired because, why not. I've seen contractors infect entire fab's in the chip industry because they plugged an infected USB stick into a 30 year old tool that hadn't seen an update in over 20. Then when the fab came back up, they did it again the next day.
Ultimately, Mythos is just another tool in the toolbox. It's great to find new vulns but it is incredibly short sighted to think it will move the needle in any meaningful way in the security industry.
I've seen things you people wouldn't believe. Attack ships on fire off the shoulder of Orion. I watched C-beams glitter in the dark near the Tannhäuser Gate. All those moments will be lost in time, like tears in rain. Time to die.
Wait, is this quoting Blade Runner to describe a real Mythos attack, or just vibing with the dread? "C-beams glitter near the Tannhäuser Gate" isn't a CVE. What really happened to your systems, and when?
Mythos actually does change that calculus. Going forward, with access to a mythos caliber llm actors are not tied to bad configs or lazy admins for access. I get that the bs is real. But it's important for you to not rest on your laurels having recognizing that salesmen sell. You actually have to pay attention to and understand the new developments your field. It's sad that the marketing department odd doing a better job than you in that manner
Mythos finds exploits largely by reading source code.
Your open source dependencies may need to be version bumped quickly, but most companies are not going to be immediately exploitable without a large scale source code leak, and an attacker motivated to spend large amounts of money/compute on finding lucrative exploits (not just any exploits).
To me the reaction has been way overblown, though again, very real for large scale open source projects.
And going forward there's not going to be as many issues due to using models defensively, e.g. this vulnerability spike is likely a one time event.
We already are using software that is ancient, with many vulnerabilities that are already in the public, we already use insecure software more than we care to admit, if Mythos is gonna help with that, it's gonna make finding (not discovering) these vulnerabilities easier because it already has the knowledge, but the enough intellect to come up with new ones.
Same applies for other LLMs
Forget whether it is Mythos or GPT 5.6, or any other specific model. SOTA models have tool likely have the knowledge and capability to create zero days from nearly every discovered and many undiscovered vulnerabilities. In the wrong hands can deploy and generate malware and submarine code that would go undetected behind secured systems. Add in the ability to clone voices, create mass social engineering campaigns.
Yet "Just another tool in the toolbox." I mean, that's not wrong!
- June 1st 2026: Anthropic files S-1 paperwork with SEC to get ready for IPO
- June 2nd 2026: Anthropic annouces expanding "Project Glasswing" to let people use their new model to enhance security of existing systems
- June 9th 2026: Anthropic releases Mythos model
- June 12th 2026: Model gets export regulations placed on it by US Gov
- June 26th 2026: US gov announces they will let some companies use new model
- August 2026: Anthropic goes IPO
The timing of all of this just seems to be a play to pump the stock. The reality is that in six months GLM-5.3 will be released open source with comparable functionality to their Mythos model. They are trying to cash in before that happens.
I would not be surprised if the US government, the people pulling the strings who actually put the export announcements onto Anthropic, actually have purchased stock in the company to artificially pump up the stock, I would bet money on it.
Anthropic losing their ability to release new models to most customers (and thereby revenue, and thereby ability to train new models) makes you think investors will value it more highly than if they could release new models to everyone who wanted to pay them?
Since training/inference/datacenters are a money sink (as you can read in any financial insights of anthropic, openai, etc.) having more customers might actually be detrimental.
Just look at the consumer side: the current attitude of most people is they'd rather not pay the actual cost of the LLM they're using. Therefore the big money is probably in an IPO by boosting your product to be so unfathomably potent, it must be ridiculously valuable to own and control.
It also helps to pretend it's actually too dangerous for the general public: high-paying government contracts only please.
Minor nitpick: S-1 filing isn't "getting ready" for IPO, it's the formal start of the roadshow clock, typically 3-6 months before listing. So August is actually on the fast side, not slow. Still, the whole stack of coincidences is oddly tight.
Nah I spoke to a security researcher who still has access to Mythos. He says it is significantly better than their earlier models for security research. Based on my one-day use of Fable that was also a noticeable step up for coding.
There's absolutely no way Anthropic engineered this to bump their IPO price. That's lunatic conspiracy theory territory.
> I would not be surprised if the US government, the people pulling the strings who actually put the export announcements onto Anthropic, actually have purchased stock in the company to artificially pump up the stock, I would bet money on it.
The same US government that labelled Anthropic as a supply chain risk? This is the most ridiculous idea I've heard all week.
Ha yeah I totally agree, that's actually one of the future posts I have in draft : the other downfall of GenAI in cyber. You can't outsource learning, and a lot of learning opportunities in the Cybersecurity industry are getting totally ruined by llms (ctf, low hanging fruit bug bounties, foss software getting burnout by AI slop and closing the gate to potential newbie willing to get involved, etc.)
Find some pretty good vulnerabilities, and at a very fast speed--one or two weeks ago, mimo-v2.5-pro(some where between v4 flash and pro), released ultra-speed version with 1000 tokens/s. gpt-5.6 sol also has a nominal 750 tokens/s
With so much cloud being at risk from AI now, soon or in the future, it seems like self-hosting or at least managed custody of your own gear is going to become more of a thing.
the idea is that you don't have half of all companies all share the same liability of a single cloud provider who is half-arsing it because of their monopoly position.
Saw this exact playbook with Stuxnet in 2010, everyone panicked, budgets tripled, then it normalized. Mythos just moves faster. Least privilege and patch cadence stopped most of this in 2005 against SQL injection, still does now.
> since they have values more compatible with western democracy.
I'm mostly investing in China these days too. Remember that they have no soul and only do this as long as it is beneficial to them to appear that way (liberal, democratic). Same for the US. This is how the game is played.
Companies have never secured their stuff and it's not because they didn't have access to Mythos. No one cares and breaches don't cost them money or customers. If I sound cynical it's because I am.
There's no functional difference between
"Hey npm says this is vulnerable, we need to fix it!" /
"Nah, later."
and
"Hey Mythos says this is vulnerable, we need to fix it!" /
"Nah, later."
I tend to agree but open weight model seem to still be lagging behind in terms of capacity, even the recent ones like GLM 5.2. If anything I hope the sudden, unpredictable changes of policy will make EU companies think twice before putting all their eggs in the same AI vendors's basket, all US based. Vendors coming back on their retention policies like they did with Fable 5 or plainly cutting the service without notice should be a gigantic red flag about your business continuity.
It's maddening how the corporate world can get shy of using any of those Chinese models, just because they are Chinese. This kind of FUD makes little sense when the inference is done in-house or by an EU/US cloud provider.
> I tend to agree but open weight model seem to still be lagging behind in terms of capacity, even the recent ones like GLM 5.2.
Haven't they been mere months behind frontier for year(s) now? And if US frontier models are going to be restricted by the US gov from a massive share of their worldwide potential customer base going forward, that also correspondingly cuts US labs’ revenue and ability to train new models, so unless Chinese models are wholly dependent on distilling more powerful models…
It's one thing to forget an Authorize attribute. That's a coaching event and procedure update. It's another altogether to not be able to see a dormant use-after-free bug because your brain can't hold the entire codebase and product roadmap at once. You can't coach a human developer on that. We all miss things this deep in the rabbit hole. The 2nd best option is to avoid this space of possibilities altogether.