International Monetary Fund Managing Director Kristalina Georgieva has expressed concerns about cybersecurity risks posed by advanced artificial intelligence models, calling for stronger guardrails to protect financial stability. Speaking on CBS News' "Face the Nation," Georgieva specifically mentioned risks associated with Anthropic's Claude Mythos AI model and emphasized the need for key institutions to collaborate on managing these threats.

The IMF chief's comments come as financial institutions increasingly adopt AI technologies. The Bank of Canada has reportedly met with major lenders to discuss AI cybersecurity risks, while Wall Street banks are testing various AI systems as U.S. regulators encourage such evaluations.

Meanwhile, cybersecurity experts at the RSA Conference 2026 highlighted significant gaps in AI agent security across enterprises. According to survey data, 79% of organizations already use AI agents, but only 14.4% have full security approval for their agent deployments. A separate survey found that only 26% of organizations have AI governance policies in place.

In response to these security challenges, technology companies have developed new architectural approaches. Anthropic launched its Managed Agents system in April, which separates AI reasoning from code execution and isolates credentials from the execution environment. Nvidia released NemoClaw in March, taking a different approach by wrapping agents in multiple security layers while maintaining integrated monitoring.

Security researchers note that traditional AI agent deployments often place credentials and execution code in the same environment, creating vulnerability to prompt injection attacks. The new architectures represent different strategies for addressing these risks, with Anthropic focusing on structural separation and Nvidia emphasizing containment and monitoring.

The developments reflect growing recognition that AI security requires specialized approaches as these systems become more prevalent in enterprise environments. Cybersecurity leaders emphasized that securing AI agents requires continuous verification of actions rather than simple authentication, comparing the challenge to managing highly privileged users with autonomous capabilities.