It is possible that an artificial intelligence capable of turning the world upside down and causing chaos worldwide is already in the hands of some people. The point is that both Anthropic, for a few weeks now, and OpenAI, for a few days, already have respective models – Claude Mythos in the first case and GPT-5.5 in the second – capable of compromising operating systems and cybersecurity barriers in a way as astonishing as it is dangerous.
Read more Stabbed and buried in a house, the cruel end of the witness to the CAM crime
First, it was Anthropic that explained that its Mythos model would have restricted access, because in the initial tests it was subjected to, it identified hundreds of critical cybersecurity flaws and vulnerable points in every single major operating system and browser we use daily.
Technology companies such as Microsoft, Apple, Google, Nvidia, and Amazon joined the coalition to explore the dangerous capabilities of Mythos and seek solutions before it falls into the wrong hands.
This is serious, but Sam Altman, CEO of OpenAI, stated via a tweet on the social network X that Anthropic’s move was “fear-based marketing.” In his opinion, that strategy was like saying: “We have built a bomb. We are about to throw it at your head. We will sell you an air-raid shelter for 100 million dollars.”
But OpenAI presented last week its GPT-5.5 model, which runs on ChatGPT and turns out to also have, like Mythos, a great capacity to find loopholes in cybersecurity that make it dangerous.
Sam Altman restricts his Cyber version after criticizing Anthropic’s “fear-based marketing”
The GPT-5.5 version with those capabilities is called Cyber and, paradoxically, OpenAI will only let a selected group of users who are “critical cyber defenders” try it.
Read more Rapes and attempts to kill a woman he met on an app: he already had a criminal record in the U.S.
The information provided by OpenAI is that Cyber can perform tasks such as penetration testing, identification and exploitation of vulnerabilities, and reverse engineering of malware. It is like a toolkit for companies to test their defenses and find security holes.
However, the UK’s AI Security Institute (AISI) has tested the GPT-5.5 version released to the public last week and its conclusions are that it is at the performance level of Anthropic’s model. “It reached a similar performance level in our cyber assessments” to Mythos Preview, which this group was authorized to evaluate last month.
Also read
The AISI is an institute specialized in the analysis of web exploitation tools and cryptography. In the highest-level testing tasks, GPT-5.5 scored an average of 71.4%, slightly higher than the 68.6% achieved by Mythos.
In a particularly difficult test, which consisted of simulating a 32-step data extraction attack on a corporate network, GPT-5.5 succeeded in 3 of the 10 attempts it made, compared to 2 out of 10 that Mythos managed. No previous model had been able to do so even once. Models still fail in a test that simulates disrupting the control software of a power plant. For now.
Read more The dollar: weakness without surprise
