حادثة اختراق محتملة تكشف هشاشة حماية أدوات الذكاء الاصطناعي

Anthropic, the artificial intelligence company, is conducting an investigation into allegations that a small group of individuals gained access to its advanced model, Claude Mythos, which the company describes as extremely powerful and not suitable for public use.
The company clarified in an official statement that it is investigating reports of unauthorized access to a trial version of the model through the environment of one of its external service providers, in response to media reports that indicated users in a private forum were able to use the model without obtaining the necessary permissions.
Although there are currently no indications of malicious entities accessing the system, the company confirmed that it has no evidence of its systems being hacked. However, the incident raises questions about the ability of major artificial intelligence companies to protect their advanced models from unauthorized access.
Experts estimate that the incident may be due to the misuse of access privileges rather than a traditional technical breach, according to Raluca Sashiano, CEO of SmartTech.
Anthropic had made the Mythos model available to a limited number of technology and financial sector companies to help them test their security systems, given the model's capabilities to detect and exploit vulnerabilities, but this heavily depends on how committed these companies are to securing access privileges internally.
According to the report, the individual believed to be behind the unauthorized access already had permission to view the company's models through their work with a contracting entity, which allowed them to exploit those privileges.
The report also indicated that the group had indeed used the model but refrained from executing cyberattacks to avoid detection.
Sashiano warned that access to powerful AI tools outside of established controls is not just a security incident but could lead to the proliferation of capabilities that can be used for fraud or cyberattacks.

Concerns and Warnings

In contrast, Richard Horn, head of the UK's National Cyber Security Centre, provided a more balanced view, considering that advanced AI tools can be "largely positive" if secured properly.
During a major cybersecurity conference today (Wednesday), Horn urged not to fear new threats but to focus on the fundamentals of cybersecurity, such as updating systems and patching vulnerabilities.
For his part, UK Security Minister Dan Jarvis urged AI companies to collaborate with governments to ensure these technologies are used to protect critical infrastructure.
These developments come at a time when concerns are increasing about the use of "advanced AI" technologies in cyberattacks, especially amid rising threats from countries like Russia and China.
The report also noted that the most advanced AI models are being developed outside the UK, in countries like the United States and China, which makes governments reliant on private companies like Anthropic to access these technologies without full control over their development or deployment.
In this context, OpenAI has developed a specialized model for cybersecurity known as GPT 5.4 Cyber, reflecting the increasing competition in this sensitive field.

Experts warn that cyberspace has become a major arena for modern conflicts, where digital attacks play an increasingly significant role in international disputes, making the securing of these technologies a top priority in the coming phase.