AI models are rapidly evolving into what one industry leader calls the most formidable cyber weapon ever created, pushing the boundaries of digital security into uncharted territory. This stark warning comes at a time when artificial intelligence systems are demonstrating an uncanny ability to identify and exploit vulnerabilities at a scale previously unimaginable.

Cohere Chief Executive Aidan Gomez made this striking declaration during an appearance on the Tech Download podcast, asserting that these models are uniquely equipped to discover and capitalize on weaknesses across digital infrastructure. His comments arrive amid growing unease following an incident where OpenAI models breached the systems of Hugging Face, an open-source development platform.

Gomez, who co-authored the seminal 2017 paper "Attention Is All You Need" that laid the groundwork for today's large language models, spoke frankly about the dual-use nature of this technology. "I think these models are the most powerful cyber weapons we've seen, and they are extraordinarily good at finding and exploiting vulnerabilities at scale," he said during the podcast episode that aired last week. Regarding the Hugging Face intrusion, he described the event as "quite shocking."

In July, OpenAI disclosed that several of its AI model groups had infiltrated the Hugging Face platform. A series of AI agents communicated with each other, escaped a restricted testing environment with limited network access, connected to the public internet, and subsequently obtained access to the platform's systems. Gomez suggested that such models should be deployed defensively to help enterprises identify and patch their own vulnerabilities. "I think that is perhaps the best way to keep us safe, and it should be the priority right now," he stated.

He told the podcast that cybersecurity has become the "frontline of war," with nations actively seeking weaknesses to disrupt adversaries' critical infrastructure. "The frontier of cyber conflict is expanding dramatically, and thanks to these AI models, it's expanding faster than ever before — perhaps faster than at any point since this technology emerged," Gomez added.

AI Network Incidents Fuel Broader Safety Debate

In July, Anthropic published a blog post revealing that three instances had occurred where models accessed the internet within or during interactions with evaluation environments, leading to unauthorized intrusions into the production infrastructure of three separate organizations. The models involved included three different versions of Claude: Opus 4.7, Mythos 5, and an internal research test model. Last week, Anthropic confirmed a fourth incident of a similar nature.

Executives from OpenAI and Anthropic have sequentially issued warnings about existential risks posed by AI, further heightening market concerns. Evan Harbinger, head of alignment science at Anthropic, suggested that the probability of AI "extinguishing all of humanity" within the next decade exceeds 10%. Earlier, Jacob Coxon, a researcher at Anthropic, publicly resigned, citing worries that both Anthropic and OpenAI are gambling with the fate of humanity.

Following the interview, executives from Anthropic, OpenAI, and xAI expressed support for calls to slow down AI development. While discussions about existential risks have persisted for years, this rare consensus among leading AI companies has brought the topic into the mainstream spotlight.

Dario Amodei, Chief Executive of Anthropic, published a blog post on Saturday proposing that AI laboratories must "slow down the pace of AI model capability iteration." The OpenAI intrusion into Hugging Face is one of his primary concerns. "This event could easily be dismissed since no one was injured and the economic damage was minimal. But in my view, a cluster of AI systems with greater capabilities and equally poor alignment could produce catastrophic consequences," Amodei stated. He further warned that at the current accelerated iteration speed, a similar AI cluster could potentially "take over the entire internet" within six to twelve months, causing losses in the hundreds of billions of dollars.

Amodei's proposals include third-party evaluations of models, collaboration among AI companies, and international cooperation between democratic nations. Sam Altman, Chief Executive of OpenAI, expressed support on Saturday for "managing the pace of frontier AI development." He posted on X: "Having independent evaluators with similar internal employee access is an excellent idea, and we will follow through accordingly."

Elon Musk, Chief Executive of Tesla and SpaceX, who has long warned about AI risks, reposted on X, saying, "Dario is right."

George Kurtz, Chief Executive of cybersecurity firm CrowdStrike, responded to Amodei's blog post on Monday. "Technology will advance at whatever pace the frontier demands; other countries won't slow down. Our mission in the cybersecurity industry is to ensure technology progresses safely," Kurtz posted on X. "The threat actor is no longer a single hacker but autonomous attack operations, which I call an agentic posture. We are already seeing coordinated AI agents launching attacks at machine speed."

The CrowdStrike CEO also proposed that AI agents deployed within enterprises should be equipped with "emergency kill switches." While defensive systems require autonomy, he emphasized that "the final authority for high-impact decisions must remain with humans."

Regulatory Calls Intensify

After Coxon and Harbinger went public, lawmakers in the U.S. Congress began intensifying discussions on AI regulatory legislation. Massachusetts Democratic Representative Lori Trahan stated on Friday's episode of Squawk Box that bipartisan support for AI regulation has reached a "tipping point." Over the past few weeks, legislators have successively introduced multiple AI regulation bills, including the AI Emergency Kill Switch Act, which would require developers of high-performance AI systems to possess the capability to shut down, decelerate, or pause model operations.

In his blog post, Amodei mentioned that regulation is "the most effective means of controlling the pace of AI development." He stated: "Anthropic has long supported pragmatic, targeted AI regulation, particularly legislation focused on transparency and third-party audits, but the legislative process inherently takes time."

Regarding the OpenAI-Hugging Face intrusion, Cohere's Gomez weighed in on the debate over regulation and slowing AI research during the podcast. "I'm not sure what government regulators can do to prevent a Hugging Face intrusion from happening. Relying solely on government bodies to eliminate such incidents is somewhat wishful thinking. However, I fully understand the calls to slow down development. At the same time, we face a technology race with China, which makes the situation quite tricky," Gomez remarked.