The people who want to be dismissive are going to find reasons, but the security experts I follow are taking the AI-based security threat very seriously. Some guy wants to make people hate...
The people who want to be dismissive are going to find reasons, but the security experts I follow are taking the AI-based security threat very seriously.
Some guy wants to make people hate Anthropic and I don't really see the point of writing a long takedown about it. I don't think we need to take a position on how much better Mythos is, because it doesn't really matter. There are more high-quality security bugs being found through a variety of AI-enabled means. The people who maintain important systems have a lot more work to do lately.
If Mythos could do what's claimed, why do none of Anthropic's numbers demonstrate that capability? Why did they need to lie so plainly? If bugs are being found through AI-enabled means, why isn't...
If Mythos could do what's claimed, why do none of Anthropic's numbers demonstrate that capability? Why did they need to lie so plainly? If bugs are being found through AI-enabled means, why isn't Anthropic presenting that data instead?
Anthropic's false and distorted numbers are worth discussing whether people are just being haters or not. Much of their claims and the expectations of security experts are built solely on those numbers which are horrifically misleading at best.
Companies are taking the Mythos security threat seriously. Companies are finding and fixing security bugs. This is probably a good thing. Could they have found most of the bugs using a cheaper...
Companies are taking the Mythos security threat seriously.
Companies are finding and fixing security bugs.
This is probably a good thing.
Could they have found most of the bugs using a cheaper existing model? Who cares?
Is the $100 million worth of free tokens pure marketing genius? Who cares?
Based on what? Their own very-limited experience with Mythos so far? Or the misleading numbers and marketing that Anthropic put out? Nobody who was willing to commit to that publicly. The only one...
Companies are taking the Mythos security threat seriously.
Based on what? Their own very-limited experience with Mythos so far? Or the misleading numbers and marketing that Anthropic put out?
Companies are finding and fixing security bugs.
Nobody who was willing to commit to that publicly. The only one who somewhat has, Mozilla, is somewhat dispelled in this article as the numbers were inflated and none represented actionable real-world exploits as claimed.
Could they have found most of the bugs using a cheaper existing model? Who cares?
We should all care about being blatantly, openly lied to.
This is a citation-heavy teardown of basically every claim Anthropic made about Mythos. The key takeaway for me was that Mythos is not any sort of generational improvement. The numbers have been...
This is a citation-heavy teardown of basically every claim Anthropic made about Mythos. The key takeaway for me was that Mythos is not any sort of generational improvement. The numbers have been heavily fudged and their methodology obfuscated to cover the fact that even Sonnet models can go toe-to-toe with it when you aren't counting single issues multiple times, with those single issues being in highly contrived unrealistic environments (again) contrary to what was claimed.
It probably isn't surprising, but since 2019's GPT-2 the "too dangerous to publicly release" narrative still falls short of the marketing.
These findings, combined with earlier results on synthetic imagery, audio, and video, imply that technologies are reducing the cost of generating fake content and waging disinformation campaigns. The public at large will need to become more skeptical of text they find online, just as the “deep fakes(opens in a new window)” phenomenon calls for more skepticism about images.
Today, malicious actors—some of which are political in nature—have already begun to target the shared online commons, using things like(opens in a new window) “robotic tools, fake accounts and dedicated teams to troll individuals with hateful commentary or smears that make them afraid to speak, or difficult to be heard or believed.”
Selling access prevented none of that, and GPT-2 wasn't the inflection point for that. Even in 2026, current spam and propaganda on the internet still very often gets along just fine with non-AI...
Selling access prevented none of that, and GPT-2 wasn't the inflection point for that. Even in 2026, current spam and propaganda on the internet still very often gets along just fine with non-AI bots with standard templates and character substitutions vs human-run social accounts spewing set talking points, occasionally with an AI-generated image or comic for extra punch. The viral Facebook BS, SEO spam sites targeting every niche, and LinkedIn post economies have been revolutionized, though.
GPT-2 was ultimately near enough trivial in compute and dataset. Maybe worse than existing methods of harm they identified. They were testing staged releases and delays to collect more usage data,...
We are aware that some researchers have the technical capacity to reproduce and open source our results. We believe our release strategy limits the initial set of organizations who may choose to do this.
While the misuse risk of 345M is higher than that of 117M, we believe it is substantially lower than that of 1.5B, and we believe that training systems of similar capability to GPT‑2‑345M is well within the reach of many actors already; this evolving replication landscape has informed our decision-making about what is appropriate to release.
GPT-2 was ultimately near enough trivial in compute and dataset. Maybe worse than existing methods of harm they identified.
They were testing staged releases and delays to collect more usage data, IMO. As it turns out they were already studying RLHF.
And they began selling access to much more powerful models.
Just felt like a big joke a year or two later when I understood it better.
The irony of this piece is that you can play count the LLM writing tropes in it. If you make it a drinking game you'll pass out before the end. My takeaway: what's the point? Mythos will be...
The irony of this piece is that you can play count the LLM writing tropes in it. If you make it a drinking game you'll pass out before the end.
My takeaway: what's the point? Mythos will be released publicly at some point and everyone can check for themselves. Until then, outsider speculation isn't adding anything useful to the conversation.
But if we're speculating... Based on how the technology works, and the history of model releases, most likely it'll be a sonnet to opus level of improvement rather than a game changer.
But the enhanced ability to turn bugs into working exploits likely really does justify caution. Which is not to say marketing wasn't a primary consideration.
As an aside, AI model naming is, I feel, really silly. Like you cannot go from Haiku and Sonnet to "Mythos" and not feel like you're maybe overselling things a bit.
As an aside, AI model naming is, I feel, really silly. Like you cannot go from Haiku and Sonnet to "Mythos" and not feel like you're maybe overselling things a bit.
The people who want to be dismissive are going to find reasons, but the security experts I follow are taking the AI-based security threat very seriously.
Some guy wants to make people hate Anthropic and I don't really see the point of writing a long takedown about it. I don't think we need to take a position on how much better Mythos is, because it doesn't really matter. There are more high-quality security bugs being found through a variety of AI-enabled means. The people who maintain important systems have a lot more work to do lately.
If Mythos could do what's claimed, why do none of Anthropic's numbers demonstrate that capability? Why did they need to lie so plainly? If bugs are being found through AI-enabled means, why isn't Anthropic presenting that data instead?
Anthropic's false and distorted numbers are worth discussing whether people are just being haters or not. Much of their claims and the expectations of security experts are built solely on those numbers which are horrifically misleading at best.
Could they have found most of the bugs using a cheaper existing model? Who cares?
Is the $100 million worth of free tokens pure marketing genius? Who cares?
Based on what? Their own very-limited experience with Mythos so far? Or the misleading numbers and marketing that Anthropic put out?
Nobody who was willing to commit to that publicly. The only one who somewhat has, Mozilla, is somewhat dispelled in this article as the numbers were inflated and none represented actionable real-world exploits as claimed.
We should all care about being blatantly, openly lied to.
This is a citation-heavy teardown of basically every claim Anthropic made about Mythos. The key takeaway for me was that Mythos is not any sort of generational improvement. The numbers have been heavily fudged and their methodology obfuscated to cover the fact that even Sonnet models can go toe-to-toe with it when you aren't counting single issues multiple times, with those single issues being in highly contrived unrealistic environments (again) contrary to what was claimed.
It probably isn't surprising, but since 2019's GPT-2 the "too dangerous to publicly release" narrative still falls short of the marketing.
From OpenAI's 2019 announcement about GPT-2:
Seems to me that holds up well?
Selling access prevented none of that, and GPT-2 wasn't the inflection point for that. Even in 2026, current spam and propaganda on the internet still very often gets along just fine with non-AI bots with standard templates and character substitutions vs human-run social accounts spewing set talking points, occasionally with an AI-generated image or comic for extra punch. The viral Facebook BS, SEO spam sites targeting every niche, and LinkedIn post economies have been revolutionized, though.
GPT-2 was ultimately near enough trivial in compute and dataset. Maybe worse than existing methods of harm they identified.
They were testing staged releases and delays to collect more usage data, IMO. As it turns out they were already studying RLHF.
And they began selling access to much more powerful models.
Just felt like a big joke a year or two later when I understood it better.
The irony of this piece is that you can play count the LLM writing tropes in it. If you make it a drinking game you'll pass out before the end.
My takeaway: what's the point? Mythos will be released publicly at some point and everyone can check for themselves. Until then, outsider speculation isn't adding anything useful to the conversation.
But if we're speculating... Based on how the technology works, and the history of model releases, most likely it'll be a sonnet to opus level of improvement rather than a game changer.
But the enhanced ability to turn bugs into working exploits likely really does justify caution. Which is not to say marketing wasn't a primary consideration.
As an aside, AI model naming is, I feel, really silly. Like you cannot go from Haiku and Sonnet to "Mythos" and not feel like you're maybe overselling things a bit.
Anthropic: Limerick gains capability to impair human respiratory system, too dangerous to release