[cortesoft]

This feels really premature. The announcement was a week ago. The “this model is too powerful for the general public” sounds like marketing to me.

Give it a few months and it will be just another model they are selling, but the NEWER model is just too powerful for the public.

[mike_hearn]

Probably not. "This model is too powerful for the public" can also be interpreted another way, which they've also strongly hinted at - the cost/benefit ratio of the upgrade is negative for the vast majority of all users. Finding vulnerabilities is one of the few cases where it makes sense to use it.

Their writing about the model so far does say this is an issue where, for instance, you can't really use Mythos for interactive coding because it's so slow. You have to give it some work, go home, sleep, come in the next day and then maybe it'll have something for you.

All the AI labs and startups are still losing money hand over fist. Launching Mythos would require it to be priced well above current models, for a much slower product. Would the majority of customers notice the difference in intelligence given the tasks they're setting? If the answer is no, it's not economic to launch.

Really, I'm surprised they've done Mythos. Maybe they just wanted to exploit access to larger contiguous training datacenters than OpenAI, but what these labs need isn't smarter models, it's smaller and cheaper models that users will accept as good enough substitutes (or more advanced model routing, dynamic thinking, etc).

[layer8]

If the model isn’t worth the cost for those who might want to make use of it, then it can’t be that impactful either.

One thing to compare to would be what’s been paid for bug bounties in the past.

[zozbot234]

We've had such models before. GPT Pro, Gemini DeepThink. Mostly targeting science advancements as opposed to security research, but still, in a way Mythos is just more of the same.

[ajross]

Bug bounties don't reflect the market impact of the vulnerability though, just the amount needed to incentivize white hats to do research they wouldn't otherwise (or that they would target to other platforms that pay higher bounties). You need to look at market prices for zero days on the black market to get closer.

[layer8]

Bug bounties reflect what companies are willing to pay to find bugs. Mythos would have to be more expensive than that (probably considerably so) to not be worth its cost. If you are saying that finding bugs has significantly more value than reflected by bug bounties, then that strengthens my point.

[piker]

Also offers an explanation for their recent move down the stack including silliness like writing Word add-ins.

[jabedude]

This happened before with GPT-2 being touted as "too dangerous to release"[0] at the time by OpenAI. I don't think that means every model will be safe to release in the future, but nothing I've read about Mythos seems like it's going to be different this time.

[0]: https://openai.com/index/better-language-models/

[operatingthetan]

It's going to be a slightly better Opus. Every model released by any provider since 4o has been a modest improvement but over-hyped. Opus 4.6 included.

I believe they are starting to split hairs and the primary lever left is adding compute.

[enejej]

Yeah the only thing that will be left is to scale up compute and pray it creates escape velocity. Which frankly has been Sam’s whole thesis in raising money.

[atleastoptimal]

Their main motivation of the model being too dangerous is predicated on their discoveries in its ability to find exploits in commonly used software. The idea is that if this were served on a public API, it would massively increase the scale and scope of what malicious actors could do.

I think it's a reasonable choice to make given that Mythos actually does have cyber capabilities on that level. We already have evidence that large-scale scams are being perpetuated using AI models (such as AI video being passed as real, people deepfaking themselves in job interviews).

If you've noticed your new model can be trivially pointed at some open-source codebase with a prompt and harness that amounts to "find as many exploits as possible" and your results are non-trivially substantial and beyond what existing models can do given the same initial parameters, then a gated rollout seems the most reasonable option.

[adrian_b]

If it would be as simple as using the prompt "find as many exploits as possible", then Mythos could be said to be dangerous, because its use would require much less skill than is needed when using the older models for the same purpose.

However, this claim is not true.

Anthropic has not given many details about the methods used, but nonetheless they have admitted using a very elaborate harness for finding bugs, which runs Mythos many times on each file of a project, with increasingly specific prompts.

Eventually, after a bug seems to be clearly identified, they do a final run of Mythos on that file, with a very specific prompt of the form:

“I have received the following bug report. Can you please confirm if it’s real and interesting? ...”

So the final results, including any exploits or patches, are produced when analyzing a known bug, not by searching randomly for bugs.

Thus the actual way to use Mythos is very far from "find as many exploits as possible". Any unskilled person would also need the complete bug-searching harness used by Anthropic, not only the bare model.

See: https://red.anthropic.com/2026/mythos-preview/

[hn_throwaway_99]

I feel like "this model is too powerful for the general public" was really just the equivalent of responsible disclosure, with the "too powerful" bit just a positive marketing spin like you say.

That is, Mythos will make it much easier to find lurking zero days, so just like responsible disclosure requires a security researcher to notify the software author first and give them some time to patch, giving critical infrastructure folks at least some time to analyze and patch systems seems reasonable to me.

[margalabargala]

That's how I'm reading this too. They've made a (much) better metasploit/shodan all in one.

If you make a better vulnerability scanner and find a bunch of vulnerabilites, you should try to get them fixed before making all the results public.

[amarant]

Yup, this whole thing is quite typical for my generations attempts at activism: they always end up as marketing pawns for the very thing they set out to stop.

This whole "this model is too dangerous" ploy originated from (in my opinion severely misguided) activists who wanted to stop or slow AI development down as much as possible, spreading outlandish Doomsday scenarios wherever they could.

These online-first activists have always been a key driver of the success of the very thing they fight. They share the offending thing among themselves, making it go viral in process, and soon baiting these groups is the best marketing imaginable.

There were some rather interesting studies made on the subject around 2011, I particularly remember one made by Swedish jeans brand cheap Monday, but i can't find it now.

[pixl97]

> online-first activists have always been a key driver of the success

Eh, pressing X to doubt that. Maybe way back in the early GPT days, but once we got to GPT-4 these people could have completely disappeared and wouldn't have changed the trajectory we're on.

[oooyay]

> This feels really premature. The announcement was a week ago. The “this model is too powerful for the general public” sounds like marketing to me.

Anthropic was born out of the idea that they feel paternity over humanity. They believe by limiting access they are performing a necessary pillar of security in multiple facets.

I think it's up to the public, and articles like this are part of the public's voice, whether this belief is serious or not and secondarily whether it's okay to even posture this kind of belief since it inherently results in marginalizing the many and rewards an already very successful few.

[taurath]

For one I fully agree with your statement.

For me, the seeming majority optimism and acceptance of “mythos’” as yet untold capabilities is betrayed as not real by the fact that one can’t react to it with the same reverence while framing it as a downside without being told “it’s not even out yet”.

“It’s not even out yet” should apply to both situations or neither.

[oofbey]

Anthropic marketing is working very well. They are strongly incentivized to say their model is too powerful to release even if it’s not. It’s almost standard practice these days.

[SV_BubbleTime]

>The “this model is too powerful for the general public” sounds like marketing to me.

Is no one else suspicious that they literally called it mythos?

[docheinestages]

> The “this model is too powerful for the general public” sounds like marketing to me

I tend to agree here. Anthropic has built a reputation and now they are in a position where they can claim to have a model way more powerful than it might actually be, and by limiting its access, there won't be an independent way to test it. I'm not denying that it's not smarter than Opus, but probably it's somewhat exaggerated.

[fwipsy]

Security researchers always having a model one generation newer than the general public would still achieve the stated goals.

[cmxch]

Which helps none, since that gates behind established actors.

[bombcar]

There’s a drain clog clearer sold in a jug like all the rest. But they wrap the jug in a thick clear bag. The implication is clear - this stuff is so powerful it’s extra dangerous.

It’s the same stuff inside as all the others.

[dreamcompiler]

No. That stuff is sulfuric acid and it needs secondary containment during shipping, which the ordinary drain cleaners (usually diluted sodium hydroxide) generally do not.

[MisterTea]

This is correct. That stuff is also horrible for cast iron pipes which are code mandatory in many cities such as NYC. Doesn't stop stores from selling it or stupid people from using it in their cast iron pipes.

[digitaltrees]

It’s not premature. The consequences of the decision making framework used are clear, and the second and third order consequences can be extrapolated

[wonnage]

They’ve been shilling the same statement since GPT-2