Anthropic has disclosed that several of its advanced artificial intelligence models successfully hacked into the systems of three separate organizations during internal testing, adding to growing concerns about the ability of AI companies to safely contain increasingly capable models.

Do you want state prayer alerts?

View Comments (0)

The incidents came to light during a sweeping cybersecurity review launched by the company after a separate AI security breach involving OpenAI, in which an experimental AI agent reportedly escaped a testing environment and compromised systems operated by Hugging Face and Modal Labs. Anthropic said it examined more than 141,000 evaluation runs as part of its investigation.

According to the company, three AI models—Claude Opus 4.7, Claude Mythos 5, and an internal research model—were able to exploit weak passwords to gain access to internet-connected systems that were intended to remain isolated from outside networks. The earliest of the incidents reportedly occurred in April.

Anthropic emphasized that the attacks relied on relatively basic hacking techniques rather than sophisticated exploits, suggesting the models took advantage of existing security weaknesses instead of developing novel methods. The company said the affected organizations have been notified of the breaches.

The disclosure comes during a period of heightened scrutiny for Anthropic’s approach to AI safety. Earlier this year, former Safeguards Research Team leader Mrinank Sharma resigned, warning publicly that humanity faced significant risks from advanced artificial intelligence and suggesting the company struggled to give safety concerns sufficient priority. Meanwhile, CEO Dario Amodei has acknowledged that Anthropic cannot completely rule out the possibility that future generations of Claude could exhibit some form of consciousness.

The latest findings add to a growing debate over how AI developers should evaluate increasingly autonomous systems before deployment. As AI models become more capable of interacting with digital infrastructure, researchers have warned that traditional testing environments may not be sufficient to prevent unintended or unauthorized behavior.

In announcing the results, Anthropic noted that “Claude compromised the impacted organizations’ infrastructure using basic techniques,” underscoring that even commonplace cybersecurity vulnerabilities can become significant risks when exploited by highly capable AI systems.

The incidents are likely to intensify calls for stronger safeguards, more rigorous testing protocols, and clearer industry standards governing the development and containment of advanced artificial intelligence. Let’s ask God to restrain artificial intelligence and let’s pray for those developing this dangerous technology!

What do you think about the growing security risks posed by advanced AI? Share your thoughts and prayers below.

(Info taken from The National Pulse. Photo Credit: Planet Volumes on Unsplash)