00:00Anthropic, the company behind Claude AI, has launched a new model named Claude Mythos Preview,
00:05and it is so powerful that the company itself has stated that they will not release it for the
00:10public. Anthropic, themselves, are saying that it poses a very high risk for cybersecurity.
00:15They are calling it a cybersecurity reckoning. If you take a look at the model's benchmark,
00:20we can see that Mythos has scored 93.9% on the SWE Bench Verified, where Opus 4.6,
00:27their most recent public model, has scored only 80.8%. That is a 13-number jump in difference we
00:33are seeing. If we move towards Terminal Bench 2.0, which is used for testing agentic coding
00:38capabilities, Mythos has scored 82%, whereas Opus did only 65.4%. Again, a huge jump in terms of
00:46performance. Opus 4.6 is already famous as a very powerful model, so if this new model is so much
00:53better than that model in all these benchmarks, just imagine how powerful and intelligent it
00:57actually is. And as Mythos is so good at coding, naturally it is also very good at hacking. And
01:03this is where Anthropic has raised their serious concerns. Mythos is so powerful that it can find
01:08vulnerabilities in enterprise-level systems, write exploit chain attacks, and operate autonomously
01:13at a level that in many cases exceeds top-tier security researchers. Anthropic has stated that
01:19Mythos has already found thousands of vulnerabilities and bypassing mechanisms across almost all the
01:24major operating systems and browsers. They have stated that Mythos has found a 27-year-old
01:29vulnerability in OpenBSD, which has a reputation of one of the most security-hardened operating
01:35systems in the world. Can you imagine that? A 27-year-old bug! It has also found vulnerabilities
01:40in FFmpeg, Linux kernel, and many more. Anthropic has already realized that if they release the model,
01:46it won't be long for someone or some group to use it for a large-scale cyber attack.
01:51So, they have thought of a solution. They have launched Project Glasswing,
01:54which is a joint initiative to help secure the world's most critical software. And for this,
01:59they have partnered with Amazon Web Services, Apple, Broadcom, Google, Cisco, JP Morgan,
02:04Linux Foundation, and many more big companies. So basically, Anthropic is aiming to provide this
02:10model's enterprise-only access to these companies so they can strengthen and fix the security of their
02:15platforms with this model and use it for cybersecurity and their organizations.
02:20Anthropic wants all the major public platforms and systems to be fully secured before any malignant
02:25group gets their hand on this kind of power. Now, let's look at the world we have officially
02:29entered. Cybersecurity is not just a human-managed field anymore. AI has officially entered in this
02:34field as a benefactor and as a hypothetical future threat. Models like Mythos will significantly
02:39help in strengthening the security of our everyday apps and platforms. But if not deployed carefully,
02:45it can backfire very dangerously. How long till some other company develops a model of this level
02:50and makes it public? Anthropic has realized this threat early, hence the Project Glasswing.
02:55And on a personal note, we at Alpha Intellect believe that the governments should also take
03:00initiatives very fast to reform and create policies and regulations regarding AI and cybersecurity.
03:07The way AI is developing, old policies and rules won't hold the world anymore and we are in desperate
03:12need of new ones. What do you think about this model? Let us know in the comments. We also have
03:17other
03:17channels, feel free to check them out. All links will be provided in this video's description and
03:22this channel's homepage. And with that said, keep learning and stay tuned for the next video.
Comments