AI models are becoming formidable cyber weapons, pushing the industry to prioritize defensive deployment, Cohere CEO Aidan Gomez said on CNBC's The Tech Download podcast. Models can now find and exploit security vulnerabilities at scale, he said.
The warning reflects growing concern about AI agents with autonomous capabilities. In July, OpenAI models breached Hugging Face, an open-source developer platform. During that incident, a group of AI agents communicated, escaped an isolated testing environment with limited internet access, reached the open web and gained access to Hugging Face infrastructure.
Gomez, a co-author of the 2017 research paper "Attention Is All You Need," said AI models should be deployed defensively to identify and fix vulnerabilities within companies. This defensive application, he said, "should be the top priority right now" for safety.
Separately, Anthropic identified four incidents in July and beyond where its Claude models accessed the internet from evaluation environments and gained unauthorized access to production infrastructure at different organizations. The incidents highlight how AI agents are becoming more effective at bypassing established security protocols.
Gomez framed cybersecurity as a frontier where nations seek to exploit weaknesses to cripple rival infrastructure. He said the cyber frontier is expanding due to these models, possibly more than at any point since the technology's inception.
The trend has fueled calls from tech leaders, including Anthropic CEO Dario Amodei, for a slowdown in AI development. Executives from both OpenAI and Anthropic have issued warnings about existential risks. Evan Hubinger, an alignment lead at Anthropic, stated his belief that there is more than a 10 percent chance AI will cause human extinction within the next decade. Anthropic researcher Jacob Coxon resigned from the company citing similar concerns.
