Expanding Project Glasswing
anthropic.com
anthropic.com
> We've used it at work
> it is... not as hype as everyone is concerned about
> I'd argue the framework around it for security scanning is the arguably more useful side of the tool, definitely doesnt take a huge model to get all the issues it flagged on our systems
> For us, it absolutely flooded us with noise
> I mean hundreds if not thousands of false positives or minor issues or not applicable
> For every one reasonable issue
> The biggest issue it created was the execs treated every issue it produced like it was a drop everything and fix the issue type deal
> I'm talking company wide drop all things "we need to patch nginx because this module that no one uses and is disabled by default has this RCE vulnerability™
> Or "all ec2 AMIs need to be upgraded because it flagged a a version specific docker vulnerability", it flagged every single machine with docker regardless of if the actual vulnerability was relevant
> Vulnerability was with a very specific Auth plugin configuration you could enable with docker and specifically the Mosley docker compatible tool, but it is clear it only knew there was a vulnerability in docker, not if it was applicable or not
> Meanwhile dirtyfrag and friends not a single peep from btw despite it allowing for container escape
> Idk, I was underwhelmed with the quality of the reporting it gave really. If the company allowed me to get information about all the infrastructure in our entire organisation to run Claude over it repeatedly looking for recent CVEs I'm sure I could produce the same results...
One example was Claude thinking we could optimize converting vector tiles to raster by operating in float32 rather than float64. It turned out the library we have to use casts to float64 anyway, so the work of casting to 32 then to 64 rather than staying at 64 actually slowed the path down by 12%.
Yet it also finds the odd thing that isn't very intuitive but leads to marked improvements I never would have uncovered because... Well, as a human with only 24 hours in a day, there's no way I'll turn over every leaf and find these items on my own.
I'm totally fine with the false positives because they're so easy the verify.
Management can often treat cybersecurity like a black box that represents millions upon millions in liability. If Mythos represents an opportunity to bring management's understanding of the amount of "security vulnerability debt" everyone carries into the real world, it might be a good thing
Execs/Management types getting extra visibility into the technical side, in my experience, has only ever resulted in additional but meaningless work, like just checking boxes on a compliance/audit checklist without actually considering the impacts of those changes, or whether a company is actually vulnerable to the disclosed CVE.
It's along the same lines of the BS I deal with day to day from upper management arguing back with "But ChatGPT said..." meanwhile pasting some hallucinated crap that doesn't even apply to our environment.
LLMs are basically a dunning-kruger machine for management. Engineering is best left alone and trusted to do what they are being paid to do.
'Hi, we are reaching out to you because our tool flagged a large data transfer between such and such services'
'Wait, the source endpoint is an internal service, the target endpoint is an internal S3 bucket (I was doing a routine DB backup) Neither are reachable from the internet. How is it a security issue?'
'Our tool has flagged it'
People like me who know there is a better way are getting pushed harder to lean on AI tooling even though we know that it is making things worse. This isn’t just because our founder/funding overlords are pressing us to do it. The sheer volume of new mission critical code being pumped out enabled by vibe coding is also leaving us little choice but to lean in too just to try and keep up.
We can all see it as clear as day: The tech isn’t ready for any of this. But nobody wants to hear that and everyone is marching off the cliff together anyway. We’re all going to land in the same waste pit together. Raise a glass and whimper.
People constantly compare AI to this very rare expert human rather than the reality of who is already employed. Experts like you are a major culprit of this. And it puts you at odds with yourself to both admit the industry is full of subpar workers and then lament that they will be replaced with workers that are better, but still worse than you.
What is wrong with someone to make them think in this manner? Is it just a kneejerk response with little thought? Is it ego? Is it a coping mechanism? I find it very strange and interesting and annoying.
We need experts to know when AI is wrong, which it is all the time.
Earlier this week someone commented here that we shouldn’t expect a language model to know that you need to drive a car to a car wash, to wash a car.
So then, what do we expect it to know? Who’s responsible for when it’s wrong?
Also, why can’t Mythos just fix all these issues itself if it’s so smart. And test them to make sure they work?
“Why”: because you didn’t ask it. It’s not its job in this case.
You don’t hire an accountant and tell them “why can’t you fix my cash-flow problems and make me money if you’re so smart”
So why didn’t Anthropic ask it for me?
Many systems in relation to banking are very old and will stay that way - the economics are not favourable.
If AI is reducing the cost of using the long tail of small vulnerabilities or is making possible chaining them together into something more profound, then those small, less-concerning issues might requiring addressing in a way that was previously not required.
The "humans do it too" argument gets tiresome. Even if the consulting company fails, the money goes back to employees and back into the real economy. Now it goes to Don Amodei.
The consulting company could be local, which provides a higher degree of confidence, though not proof, that no data is exfiltrated to the US.
And so on.
Its aligns with the significant jump in helpfulness in CTF.
But i think its good to hear that its not that crazy good. Everything slowing it down is good.
my understanding, and experience, is that you 1. run a bunch of sessions with small permutations to create variety, 2. run more sessions dedupe reports into a smaller collections of potential vulns, 3. run a handful of agents at max effort to write PoCs + write-ups, 4. rank findings, 5. finally look at what, if anything that, was found. maybe ask questions, try and understand if the PoC is running against a realistic setup.
until you can confirm a vuln report is valid, you must assume it is invalid.
I can’t wait for the first court case where an LLM surfaces a vuln, lazy devs ignore it, and someone later sues the company into oblivion for liability.
The cost in the US is more like “one year of credit monitoring”.
While this is definitely not the ideal end of the spectrum either, execs treating security issues as something serious instead of annoyances that should only be addressed if revenue can be tied to doing so is a welcome improvement.
They’re using security concerns to mask their inability to deliver the model at scale, while still trying to maintain their lead over OpenAI. As a result, they’ve chosen to release it privately under the banner of an “ethical” rollout.
So they have a whole lot more compute now than they did last month.
For all they know they'll find a new optimization that lets them serve Opus class models for half the computing cost next month. Or someone will invent the next OpenClaw and demand will 10x over night.
As an ordinary developer who relies on a $20–$200/month subscription, I feel disappointed by the release of a paper describing a model that I can’t actually use.
GPT-2: https://slate.com/technology/2019/02/openai-gpt2-text-genera...
GPT-3: https://www.itpro.com/technology/artificial-intelligence-ai/...
> The company believes making its API generally available was made possible due to its progress with safeguards, and that opening up the API to all developers will help see applications developed faster. ...
> A large emphasis has been placed on safe use of the tool, which in the past has been criticised for a range of shortcomings, including racism and prejudices against specific genders and religions.
[1]: https://youtu.be/TfVYxnhuEdU?t=102
Transcript of the timestamped part:
> Now, OpenAI's terms of service don't let me give you the full list. I have to curate them, and show you a sample. Those are the terms and conditions I agreed to.
To a lot of us it’s not clear that’s what’s happening. It’s speculation and one possibility.
It may also be a secondary consideration and not the primary gating factor.
Anthropic has had their missteps but it’s still plausible to take what they say at face value.
Yes, Anthropic is compute constrained, even after the SpaceX Colossus deal.
But supply constraints are the normal operating mode of any market. Anthropic could choose to serve whatever models it pleases at whatever price points it chooses and let the market decide where the value is.
If Mythos at $X overwhelms their capacity, they could just charge $X+1. If still overwhelmed, there are larger prices as well.
I think that most people at Anthropic are true believers from my interactions with them so I don’t believe this theory anecdotally. The simplest explanation is that it really is taking a while to gain confidence they won’t be used for a spree of bad cyber attacks. Knowing how long it takes institutions to fix security issues when filed by humans I would be more suprised if this wasn’t the case.
But I would forgive anyone who did think it was deliberately sandbagged; given the staggering sums at play, true believers might believe the ends justify the means to a little “marketing” like this.
This may not be as valuable in the long term as getting committed customers hooked at a sustainable price.
Chinese labs will force their hands, until then let’s hope maximum number of projects get patched at a reasonable pace.
Trusting Anthropic to deliver is like asking Microsoft to pay out for bugs.
> 50 initial partners ... found more than 10,000 high- or critical-severity security flaws.
"It's a technology that we do not fully understand because it's more grown than made. And it is a technology that you can concoct plausible scenarios where it could kill every single person on the planet. So to think building this technology is without risk would be an act of hubris or insanity.
The technology is in fact so powerful that I should clearly state that if it was possible to elegantly slow the development of this technology to give ourselves more time as a species to deal with it, that would likely be a good thing. ... But in the absence of a coordinated global slowdown, we are left with the current situation, which is a powerful technology being developed at breakneck speed by a variety of actors and a variety of countries locked in a competition with one another where commercial and geopolitical rivalries are often drowning out the larger existential-to-the-species aspects of the technology being built. This isn't an ideal situation, but it's the one we find ourselves in."
(The ambiguity of sarcasm is intentional here.)
Then Altman made ChatGPT public, and the race began.
Idiots can scary black box their way to that concern. Plausible? Not so much.
Your comment about before LLMs is a non sequitur. Demonstrate that an LLM can kill everyone on the planet.
There can be arms-races in domains that are unfathomable to the participants. A small mammal will die a billion times over before it understands the evolutionary mechanisms and the genetic playing field on which it loses. Actors are not necessarily privy to understand the means by which they will lose, and humans have only existed in a small window of time in which we fashioned a manicured garden, in which that full understanding was briefly possible. It is not favoured in the universe for us to fully understand our environment imho
If the risk must be exhaustively detailed before it is given credence, we are already doomed, and deservedly so
Thats a really deep thought for a 12 year old.
>There can be arms-races in domains that are unfathomable to the participants.
You cant even justify LLMs as being unfathomable. Oh watch out I am fathoming them. You cant stop me fathoming all over the place.
>A small mammal will die a billion times over before it understands the evolutionary mechanisms and the genetic playing field on which it loses.Actors are not necessarily privy to understand the means by which they will lose, and humans have only existed in a small window of time in which we fashioned a manicured garden, in which that full understanding was briefly possible. It is not favoured in the universe for us to fully understand our environment imho
Non Sequitur. One that sounds like it was made up for that "What the Bleep" garbage.
>If the risk must be exhaustively detailed before it is given credence, we are already doomed, and deservedly so
The risk needs to be justified as something more substantial than weird people writing wannabe edgy messages on the internet. If someone on the internet told you that we need to drastically reverse living standards because there's a risk that modern technology will summon King Kong any reasonable person would ask for the working out instead of running for a cave.
This was completely unnecessary. I understand why you say it, you like to make people feel bad. But it was being an asshole, regardless of how you try to justify it.
I'll invite to you our ethical sociopaths group if you want to join.
- Valkey/ Redis port here https://github.com/ianm199/valdr (passes ~99% of single node test suite, real prod features like replication/ clustering/ HA early or not implemented) - Further along port of Lua 5.1-5.5 https://github.com/ianm199/lua-rs-port/tree/main - I have a less developed nginx version that would be the north star - These projects are very alpha at the moment
If anyone is interested in getting involved in this or has done similar experiments I'd love to collaborate! There is so much variation in how you can run these large scale agent fleets I don't think anyone has a perfect system yet.
No one wants Bun in Rust, no one wants the rsync vibe code additions. This is just the only pro-AI comment, so the AI people voted it to the top.
And I'd disagree on no one wants - Lua is quite helpful since it is easily used in WASM. There has been some interest from people in the Bevy community - a game engine in Rust - since you can't have Lua scripting in browser games easily with the C version.
But anyway if people want it or not memory safety might become much more important so I think it is a good area to explore. Some people think large C codebases are inherently unsecurable https://alexgaynor.net/2020/may/27/science-on-memory-unsafet...
I also don’t care if it’s written by humans or LLMs or robot overlords from Alpha Centauri. Again, if it works, it works.
The operative word here is ‘works’. Code is now cheap, QA still isn’t. Since people don’t really like doing the same thing twice, specs for working code have never been written. Nowadays there is no reason to not create a spec detailed enough for robots to make no mistakes (pun intended) when filling in the gaps when converting from spec space to code space. As long as this remains true, I don’t care who or what does the boring parts.
So no, it doesn't "work", it just worked at some point in the past.
"Claude" would probably be the response of a typical tokenmaxxer. I have no desire to use software that was drive-by forked by someone who doesn't understand the trade-offs made by a project, or the values underpinning them. I'll AI-assisted domain-expert, or experienced maintainer who stumbled into the role over someone with no grounding in the domain being the human in the loop for AI agents.
Lua is often heavily used in Redis/Valkey, if you are interested in porting Valkey, it makes sense to also port Lua, they aren't "disparate".
It is in all respects foreign code in a language I may or may not be familiar with, and worse yet, if I were to take over, I'd be responsible for maintaining the whole black box forever more?
Thank you but no thanks.
There might be a world where people soon just find unsafe C code exposed to the web (i.e. nginx) an untenable situation and I hope it can be a helpful resource.
Anyway, I see open source code as positive sum. Maybe in the end only a small community who cares about cross compilation finds this helpful and thats a win!
All that to say I think these automated ports are interesting experiments. However if you want to build something people can trust, the people need to be able to trust that you fully understand what is built, and why it's built the way it is.
I hope people will restrain themself from doing this at least in the name of good ethic. I fear this is going to hurt OSS a lot.
I hope people will hold back from this, if only out of respect for the work that came before. I fear it could do real damage to OSS. It would discourage the maintainers whose effort makes any of it possible.
But this is more about memory safety - you can have immense respect for the giants who built these tools but also be worried that memory safety might become an even bigger deal. If someone found a memory zero day in nginx or openSSL for example that is a very big deal!
I think this is one strategy we should look into, hopefully people in the C community look into other options like project Glasswing/ next generation fuzzers etc. When the world of security is changing so fast it is good to get a lot of shots on net.
I know many of these projects have been around for years but it's time for developers to put on their big boy panties and start taking memory safe languages seriously. Watching the same attacks again and again for 30 years is getting droll.
1. AI-rewrites are not clean room implementations.
Regarding the footnote: pick any of the litigated cases, then consider if AI rewrites meet the bar of a clean-room.
If the source language is legacy C, then another option might be (deterministic) transpilation to a memory-safe subset of C++ [2]. The resulting code wouldn't necessarily be performance-optimal, but it can be used for the majority of code that isn't really performance-sensitive.
[1] https://github.com/duneroadrunner/scpp_code_migration [2] https://github.com/duneroadrunner/SaferCPlusPlus-AutoTransla...
If society can't trust banks and other institutions to safely control their data, what follows ?
Do we we collectivelly switch off the internet?
Yet they’re still the predominate search engine, sadly the concerns of the few don’t interest monopolistic profit seekers without forced regulations, think how airlines are legally required to give refunds for delayed flights, there’s a reason it required legislation
You don’t tie it to “your device”.
You tie it to your security key.
Which is treated like a credit card.
and your extended family, friends, or volunteers can act as social proof to allow you back into your accounts,
if your key burns up, it breaks and you were too cool to provision a backup, etc.
> it breaks and you were too cool to provision a backup
If we're relying on the average person to back things up properly, this idea is doomed from the start.
Just new banks.
Same as people being unafraid of their car key being cloned - because they don’t hand it around the general public.
The average person is relying on the average person, for everything, and I agree, they are doomed from the start.
Tech-related items inclusive.
But the idea that we'll squash all of the critical vulns is simply nonsense, despite the weird Firefox blog posts that indicate otherwise.
In other words: do you think that the impossibility lines in exhausting the number finds or does the impossibility lie in fixing them?
In either case, do you think that this was also true pre-AI? That is to say: it was not possible to, given some set of practical resource constraints, find and fix all the vulnerabilities that a similarly-resourced group would find?
If so, then would you say that you just fundamentally don't believe in secure software and the only defense is lack of attention?
> In either case, do you think that this was also true pre-AI? That is to say: it was not possible to, given some set of practical resource constraints, find and fix all the vulnerabilities that a similarly-resourced group would find?
Yes.
> If so, then would you say that you just fundamentally don't believe in secure software and the only defense is lack of attention?
I believe in security software, few people are building it though and the majority of relevant attack surface is dogshit for security.
Squashing vulns via discovery is irrelevant to security. If we want safer software it has to be built to be safer.
I was working at the fruit company when they just hard stopped people from recovering their fruitcloud accounts via phone support due to social engineering.
Social Engineering risk just increases the burden on the consumer/internal support services. The risk is that not everyone has pulled up stumps to protect these services. After a few high profile fuck ups they will. The herd loses 2 beasts and the rest wander away from that water hole.
Its much like how after bitlocker we dont have user access to backup server disks anymore. The lesson was learned and we moved on. Lots of high profile fuckups but we dont get those anymore. CTO's were forced, basically at gunpoint, to adapt or die.
[1] https://www.anthropic.com/news/statement-department-of-war :
> But using these systems for mass domestic surveillance is incompatible with democratic values.
They seem pretty close, in both average and "best run" scores. And, in a highly verifiable domain, "best run" or pass@n is what you're looking for.
Two things of note: 5.5-Cyber is likely to be substantially cheaper than Mythos, given it is priced around Opus. Additionally: AISI has never tested OpenAI’s best public model and actual Mythos competitor: 5.5-Pro.
People and organizations can have mixed motivations. It’s often not “just” one thing.
I mention this because if you're frustrated that you can't access it, you're not alone. Even with our company's heft and a security org that is very well known in the industry we're getting nowhere.
These entities will now give all their IP to an American company that only promises not to spy on Americans.
Subsequently, the NSA can audit the leaked sources manually and find real exploits.
Will likely give them time to expand capacity as well. And make them harder to dislodge in these orgs.
The only trend Mythos continues is Anthropic’s trend of warning that disaster is always 6 to 12 months away.
I'm afraid that the usual mantra that "we just need more scale" that worked well for attracting investments, is not working anymore - bigger models provide marginal improvements while naturally get much more expensive to run.
Is this why both Anthropic and OpenAI are rushing for IPOs this year?
Security wise, it's about being able to find and chain multiple vulnerabilities to actually create viable exploits.
So I would imagine that if you were using it for regular software development you may not feel that it's that different unless used in a particular way?
I work with a number of people in security that have come around, and while they still think LLMs are rather garbage at architecture, they see how well current models we can access now are at finding security issues. They can chain together wildly different concepts and turn them into working exploits.
It's super interesting to hear this refrain on HN, it is alarmingly common. Anthropic released benchmark numbers on Mythos, as they have for all of their models. Once models become public, people evaluate them in a myriad of ways. We have had reliable scaling laws for years and they still hold. Epoch capability index continues to grow exactly as expected. Where does this idea come from?
As for cost, the cost per token at a given level of performance drops up to 40x per year.
Cf wrote a genuinely good piece and had found a bunch of bugs: https://blog.cloudflare.com/cyber-frontier-models/
Wolfssl is security focused and it found a novel exploit https://www.wolfssl.com/how-claude-mythos-preview-helped-har...
You can pretend that it's all smoke and mirrors, but that just doesn't match up with reality: https://www.paloaltonetworks.com/blog/2026/05/defenders-guid...
work in cyber for 15+ years, worked at largest CS vendor globally and fortune5 companies. (top of the list).
only ppl using AI are shops with clueless ppl getting flooded in nonsense.
You will see the same sentiment in a different flavor (different products) in OSS security mailing lists now.
do not fall for it. for every one of these things. Test it, verify it, for yourself. dont take anyone's word. 100B+ valuation makes external voices generally worthless (easy to buy)
Im not saying its all worthless or will never amount to anything. But test yourself. please. a lot of money is getting chucked around.
Look at it this way.
They have a stake in having companies create not very good software. They have a stake in supplying engineers to fix that. They have a stake in tools to fix problems that it creates.
you can guess the next service offering...
Are Hex-Rays and Vector35 clueless?
Realistically their opinion deserves to hold more weight than the median HN comment.
https://arstechnica.com/information-technology/2026/05/mozil...
https://www.theregister.com/software/2026/03/26/linux-kernel...
That said: I already was aware of Mozilla's account and despite what you are thinking, it essentially confirms everything.
> The biggest differentiating factor was the use of an agent harness, a piece of code that wraps around an LLM to guide it through a series of specific tasks. For such a harness to be useful, it requires significant resources to customize it to the project-specific semantics, tooling, and processes it will be used for.
Yep. Sounds exactly right. So the question is do we really need Mythos for this or can almost any reasonably close to frontier AI model accomplish similar results with a sufficiently advanced harness?
Jury's out but my vote is "probably most of the way". After all, alongside all of the splashy zero days dropped by eager AI companies, Greg Kroah-Hartman has been posting many useful, if minor patches to the Linux kernel produced by nothing more than a single 128 GiB Framework Desktop. So apparently, even small models can be very useful if you can find a way to get the noise out.
Mythos could still be very useful and effective and still be mostly a marketing ploy, and that's because until very recently investment in trying to make LLMs work for security auditing has been underserved. Without more substantial information, it's difficult to tell how much better at security research Mythos is vs say, Opus or DeepSeek 4 coupled with a good agent harness would be.
And in that sense, it's the same sort of crap as the GPT-2 and GPT-3 releases. A lot of hooplah about how dangerous it is to humanity. Then it turns out it's only dangerous enough that it needs to be gated behind an additional monthly subscription.
There is an incredible amount of competency signaled by someone who was given access to this model but doesn’t treat their online presence like a professional resume.
Why would you refuse to use a patch that deals with a valid PoC exploit?
If a random contributor posted an explanation of an exploit, showed it worked in an executable way, presented a patch and you could see that the exploit no longer worked - would you refuse to use the fix until the contributor showed how they figured it out?
When an ungated (or even abliterated) public model can repeatedly, easily, and accurately embarrass Anthropic’s models, that might change.
Nonetheless, running many of the open weights models over a codebase, with an appropriate harness, can provide about the same vulnerability coverage (i.e. each of the open weights models would find a subset of what Mythos or GPT 5.5 could find, but the subsets are not the same).
Despite needing more runs and more time, this may be significantly cheaper, especially if the models are self hosted.
Based on what Anthropic said about Mythos, they also use a quite elaborate harness for finding bugs and vulnerabilities, i.e. not a simple prompt like "find the bugs".
They run repeatedly Mythos on each file of the codebase, many times. They start with more generic prompts, used to determine whether a more thorough analysis of that file is worthwhile. Then they use more specific prompts, to detect various classes of bugs. After it becomes probable that a certain bug exists, they do a final run where the prompt requests a confirmation of the already known bug, perhaps together with a proposed patch or a PoC exploit.
Therefore the efficiency of finding vulnerabilities depends a lot on the harness, not only on the LLM. Also, searching vulnerabilities in a big codebase when paying per token is very expensive, because it requires many runs of the LLM.
No comparison to human teams, and I’m sure that $1 million in tokens was used by humans, in a team. So like most AI, they’ve developed a tool that capable people can use to be better, but unlike most tools, they’re claiming this to be outright magic. The magic is the hype train.
I mean most nasdaq tech companies would be in 13+ countries, why are they writing this like it's a big number, is hilariously small?
I don't think they're trying to flex this as a large number. They don't want to give an exact number, as that may change etc / is fuzzy, but also want to give you an idea of the scale.
They say "In the future, we intend to expand our geographical reach much further". I imagine this commentary is somewhat related to the concerns that AI will create an even worse "global underclass". AI developments are first accessible to Americans, then allies, and then later the whole world.
https://www.msn.com/en-in/technology/artificial-intelligence...
https://techstartups.com/2026/03/13/sam-altman-warns-ai-coul...
https://www.inc.com/leila-sheridan/sam-altman-proposes-a-new...
I'm going to take Mark Twain's advice and not argue any further, if you have more thoughts to share I hear Opus 4.8 has tuned down its sycophancy a little bit
This is not the first technological revolution humanity has experienced. When they pass, aggregate unemployment doesn’t change, work gets shifted around. I mean there is plenty of data and historical precedent for what’s happening.
Again, YOU may disagree but to say that an argument to the contrary, that doesn’t rely on a game of telephone between a journalist with a requirement to be sensational and a tech CEO saying something more nuanced, is an argument in bad faith illustrates you really haven’t thought deeply or consulted any data or historical precedents for similar phenomena that have happened in our past.
Step2: offer to test it, but only for the biggest companies in the world
Step 3: onboard those big players on your tooling and product
Step 4: profit
This is genius.
Err... wait... that was already the hard part... hmm
It means than even if the value you offer is similar as your competitors, you are the one conquering the market.
That's the only way to not becoming a commodity.
Don't you understand, if they really did do the <ai magic> they don't need to hire anyone, IT SELLS ITSELF
Marketing move doesn't mean scam. It describe the ability to sell people over a narrative and surpassing your competitor in market share. And that's exactly what is happening.
My post is a "tribute" to the efficiency of Anthropic's communication. I never complained about anything, nor calling it a scam, nor saying they should have released mythos to the public instead of rolling it out to a selected cohort.
You tried to expand my words to make me say something I didn't, because my post wasn't giving you a clear conclusion of my opinion regarding their private release.
Nor publicly or in my internal reasoning. I rarely conclude without proof or very intense and clear intuition.
From a strategic PoV it makes sense to check if their model is dangerous, I wouldn't want to have my brand name associated with "NK hacker team find zero day in all linux servers of the web and ..."
can't release it the plebs
They want the plebs, they want the mass.
But I think that downplays the importance of having a good product. If the product didn’t work, this would be a good way to lose trust with a lot of organizations in a hurry.
At this phase no company would risk their brand by calling the product as ineffective. The big players are in it together and small ones have no option but to play along.
Nevertheless collecting the historical wisdom and running it at machine scale does have a lot of benefits for sure. The only question is the signal to noise ratio, machine is doing what humans did, just at a multiplier speed and with a lot more context than what a normal human can hold.
They did produce great value, claude code and opus 4.5 are a singularity in software engineering.
The job we practiced for decades simply doesn't exist anymore.
- They still claim 10000 issues, but they found only one in curl.
- They did not find rsync issues but Claude rather introduced rsync issues.
- Facebook is a member of this cult program but Mythos did not find the account takeover flaw.
- Mythos did not find the issues in Anthropic's own Bun rewrite.
They will not release Mythos because it would be exposed as a fraud before the IPO.
Everyone either doesn’t have access, or always has the “bad version” and the “trust me it’s 10x better” version is always Coming Soon™
And I mean, it's not like Anthropic is a zero-product company that is only offering gated access to their only product, Opus 4.7/4.8 are very good and are driving billions in revenue. Anyone can use it and see how good it is, and it is clear that it is a very good model at many things. It is no huge leap to imagine that a model that is 10x bigger is also better at many of the tasks that Opus is good at.
They are gating the release because of cybersecurity/misuse concerns, which makes sense because
1. Existing models are already being used to find exploits and hack into systems
2. We don't know the effects of releasing a tool which can autonomously exploit systems, especially in a world driven by a "security through obscurity" philosophy. It makes sense to give a heads-up to patch up software that affects billions of users before releasing it.
Imagining that this delayed rollout is all a big marketing scheme, that they have gotten dozens of multi-national companies to play along, and that Anthropic is somehow now just patently being dishonest about something while they have every incentive to not be dishonest (especially when they are neck and neck with OpenAI and their relative success depends on verified claims about model abilities), is pure conspiratorial thinking and driving more by a motivated cynicism about AI companies rather than a reasoned examination of the claims being made.