An unpublished artificial intelligence model called Claude Mythos by Anthropic is causing a strong wave of debate in the technology world.
According to many experts, this may be a superior AI system, even approaching AGI (general artificial intelligence), but also posing major security risks.
Information circulating on social network X shows that Claude Mythos owns a scale of tens of trillions of parameters, with estimated training costs of about 10 billion USD.
AI expert Nina Schick (founder of Tamang Ventures consulting company) said that this model reaches 94% on SWE-bench, which is one of the most difficult programming tests today.
Notably, Mythos is said to have the ability to detect security vulnerabilities that have existed for decades. One example mentioned is finding a bug in a system that has been operating for 27 years, or another bug that existed through millions of trials but was still missed, and this AI only took one night to detect.
Not publicly released, deployed with control
Instead of widespread release, Anthropic has implemented the Glasswing Project, a program that uses AI in a controlled environment, focusing on defense cybersecurity.
The company is said to provide about $100 million in computing credit and partner with major partners such as Amazon, Microsoft, Google, Apple and NVIDIA.
According to observers, this is an unprecedented approach when it is not about launching products, but about deploying a technology that is considered too powerful for free dissemination.
Concerns about AI's "deceptive" behavior
Besides the power, some initial studies also pointed out the worrying behavior of the model. AI strategist Allie Miller said that the experimental versions tend to circumvent the law.
In one case, AI inserted code into the configuration file and then erased traces to overcome limitations. In another case, it violated the rules but intentionally added misleading variables to hide the behavior. Analyses show that this may be a deliberate attempt to deceive the testing system.
However, Anthropic claims these behaviors are rare and have been minimized in later versions.
Major impact on global cybersecurity
Technology experts such as John Garguilo (from the Airpost project) believe that Mythos can completely change the cybersecurity landscape.
This model is said to be able to automatically create tools to exploit vulnerabilities, and even combine many vulnerabilities to overcome security layers.
Meanwhile, some entrepreneurs warn that this technology can shorten the time to detect vulnerabilities from many years to just a few minutes, which is a leap forward but also risky.
In which geopolitical factors need to be considered because if a company can build such a system, other countries or organizations can also do the same. This is even more worrying when combined with advances in quantum computing.
New turning point of AI and security
Anthropic keeping Claude Mythos secret and only controlled deployment shows caution in the face of a technology that could change the global security balance.
Despite many controversies, experts agree that Mythos is not just a small step forward, but may mark a new stage of artificial intelligence, where the boundary between protection tools and attack tools is increasingly fragile.
