Why Anthropic set off an alarm in the industry: AI is already worrying about its potential to hack systems

It is not the first time that a Silicon Valley heavyweight talks about a technology as a danger. In fact, OpenAI, creator of ChatGPT, had said in February 2019 that its text generator was “too dangerous” to hit the market. There was even talk of “risk of extinction.” Three years later, he launched his chatbot and took the world by storm: today, generative AI is already beginning to compete with Google in the daily life of the average user.

The increasingly interconnected world in which we move depends, ultimately and in almost all cases, on the network connectivity. From a home banking operation to a message sent on WhatsApp, the large companies that handle these services depend on how secure their systems are.

Mythos Preview, according to the company itself, already shows a unprecedented ability to detect and exploit critical flaws in software autonomously. The company claims that the system can find vulnerabilities on a scale far greater than that of an advanced human researcher, including serious errors in operating systems, browsers, widely used software and even key components of digital infrastructure.

Due to this level of risk, Anthropic decided last week to limit its access to a closed group of more than 40 technology and cybersecurity organizations, including Apple, Google, Microsoft, AWS y CrowdStrikewithin a program aimed at defensive uses.

In fact, this week Anthropic released its new version of Claude, 4.7, and the company had to clarify that it is “less risky” than the edition that generated controversy: “We are launching Opus 4.7 with safeguards that automatically detect and block requests that indicate prohibited or high-risk cybersecurity uses,” Anthropic said in a statement.

Although several companies and specialists came out to downplay the tool, an independent investigation detected a substantial improvement in the model’s offensive capabilities. The AI ​​Security Institute’s evaluation found that it can chain together complex attacks on vulnerable networks and complete, in some cases, a 32-step end-to-end corporate simulation. This, until a few years ago, was difficult to imagine with so much automation.

“The Mythos news is something that we have already been talking about and analyzing in many of our benchmarks focused on hacking. There is constant progress of the models in their offensive capabilities, and they are in a way consequences of the improvements that the models have in code development,” he said in dialogue with Clarion Nicolás Waisman, Chief Security Officer of XbowAI-automated offensive security platform.

“Companies have to seriously start thinking (and acting) about how their defenses are going to manage the progress of these models, because the scale and speed of offensive capabilities are going to have a strong impact on defense,” he added.

Clarion He also contacted Mandiant, Google’s cybersecurity company, to find out its position on this problem. “General-purpose AI models now excel in vulnerability detection, Drastically reducing the barrier to entry for malicious actors and enabling mass exploitation campaigns. This opens a critical window of risk for companies,” said company spokespersons.

Mandiant focused on the operational impact, warning that AI is already shortening the time between the discovery of a vulnerability and its exploitation. In a document published after the Mythos case, they pointed out that increasingly capable models already help detect flaws and generate functional exploits, something that lowers the entry barrier for attackers and can push massive exploitation, ransomware and extortion campaigns.

CrowdStrike, another leading player in the sector, agrees that the case goes far beyond a specific launch. In a statement, he maintained that Mythos “matters for every company” because frontier models simultaneously raise the ceiling of offense and defense, and stressed that the challenge is no longer just about the security of the model itself, but how it is controlled once it enters an organization and connects with data, internal systems and workflows.

CrowdStrike, which participates in Anthropic’s closed program, also added that in its latest report it detected a year-on-year increase of 89% in attacks by malicious actors who already use AI.

Google, for its part, points to a conclusion that is in line with the industry: companies can no longer continue to defend themselves with patching and response processes “at human speed.”

The speed with which AI operates is a tool that can be very useful in defending systems. But, like everything in cybersecurity, there is no magic solution. Everything comes with a counterbalance: if a device exists and is connected to the internet, someone is going to try to attack it.

In this hyperconnected world, the biggest threat is not a hacker or a group of attackers, but those who They already use AI to their advantage to compromise systems.

The speed of human response is no longer enough to match that of artificial intelligence, so those who defend systems will increasingly depend on AI-based tools.