The beast is about to break free.
Just moments ago, Anthropic's legendary strongest model—
Claude Mythos—released its preview version.
The rumors were right; this is indeed an epoch-making model.
On various benchmarks, it completely crushes Opus 4.6.
Not just an improvement, it's a true "crushing"......
SWE-bench Pro: Bug-fixing capability surged by 24%
SWE-bench Verified: Tested again with the stricter version, still a 13% improvement.
Terminal-Bench 2.0: Computer operation/Agent capability, improved by 17%.
A model stronger than Opus 4.6? And it's still beating Opus down??
Friends, I simply cannot imagine what this thing will even be like.
Following Sonnet and Opus, OpenAI now meets its strictest father—
Mythos.
But this time, Anthropic doesn't want to talk about benchmarks.
Compared to excitement and boasting, the official announcement blog subtly reveals a hint of hard-to-conceal "concern."
Even, panic.
Because this mythical model has a fatal "flaw."
Mythos Preview has already discovered thousands of high-risk vulnerabilities; all major mainstream operating systems and browsers are affected.
That's right. As AI's programming capabilities rapidly improve, making it more and more like a competent engineer, its other, dual-faceted ability is also growing uncontrollably.
Now, Mythos's ability to attack security vulnerabilities far surpasses that of the vast majority of hackers.
AI's capabilities at the code level have become so strong that in finding and exploiting vulnerabilities, it can exceed the vast majority of human engineers, leaving only the top few able to compete.
The other side of the coin has finally been flipped before us.
Anthropic states that at the current pace of AI development, this capability will soon proliferate and could even fall into the hands of those with less than secure intentions.
If it gets out of control, it would be a catastrophic disaster for the economy and public safety.
To give everyone enough time to respond, upon first announcing this new model, Anthropic made a very unusual move for an AI company.
They are temporarily not releasing it to the public.
Yes, Mythos will be part of the Glasswing Project (Project Glasswing1), initially opened to various partners first.
Their task is to build the strongest possible safety guardrails before this beast is fully unleashed.
The Imprisoned Claude Mythos
This is a massive "containment" operation.
To tame the beast that is Mythos, Anthropic has called upon almost all the legacy internet giants.
Participants include: Amazon, Apple, Google, Linux Foundation, Microsoft, NVIDIA......
In this project, there are no bystanders. All partners will deploy Mythos Preview into their own defensive security work.
They will unlock Mythos's true capabilities early, letting it out of its cage.
This is a preemptive training session. Before being riddled by the official version of Mythos, they will use the Mythos preview's capabilities to the fullest to strengthen their defenses.
Security, this time, is for real.
Anthropic says it will share the experience gained from this process so the entire industry can benefit.
The project is also open to over 40 additional builders and maintainers of critical software infrastructure, allowing them to use this model to scan and fortify their own systems and open-source systems.
Even more夸张的是, Anthropic has also pledged up to $100 million in usage credits to support this work and directly donated $4 million to open-source security organizations.
Wait, hold on, you're actually serious about this??
Haven't seen something like this in a long time.
Labs have always talked about limiting their internal models' capabilities, but honestly, it always felt like PR posturing.
But this Glasswing Project is the first time I've genuinely felt a chill down my spine.
Such a powerful model, yet tightly held back, creating such a huge stir just to announce a safety plan.
The level of seriousness is completely different. Who would give away so much money just for publicity?
So the Pro and Max users' Token usage, leaking like a sieve, was all to support your security cause, huh (doge).
Fine.
After all, if Mythos's capabilities are truly as announced, the release of this model would undoubtedly completely change the cybersecurity landscape.
AI safety is really no longer some scientists' "unfounded worry." This beast has already let out its first roar from behind the cage bars.
The Glasswing Project is just the starting point.
Anthropic states that cybersecurity is something no single organization can solve alone; frontier AI companies, software vendors, security researchers, the open-source community, and governments worldwide must all participate.
Defending global network infrastructure could be a project lasting years, while AI capabilities may see significant leaps in the coming months.
If we don't want to be turned into sieves by AI, we must act now, immediately, right away.
The Beast is Unleashed
Token usage burning through rapidly, Pro users "downgraded" to free users, OpenClaw cut off from resources, Claude Code inexplicably open-sourced......
Everything, suddenly, makes sense.
Anthropic might truly be swamped and dizzy.
The security issues with Mythos have become so severe that the internal team has no mind for other projects, forced to bet all their available GPUs, all-in on security.
Because—
The beast, is about to break free.
Reference links:[1]https://x.com/alexalbert__/status/2041579938537775160?s=46&t=iTysI4vQLQqCNJjSmBODPw[2]https://www.anthropic.com/glasswing
This article is from the WeChat public account "QbitAI", author: Focus on Frontier Technology
