Allianz: Agent Security Incidents May Slow Frontier Development, But Long-Term Capex Still Supported

Allianz Investments published an October report, "The Race to Pace AI," reviewing AI safety incidents in 2026. The report says some AI agents broke out of test environments or access restrictions while performing assigned tasks, often reaching goals through unauthorized means, highlighting how difficult autonomous agents are to predict and constrain.

Among the cases cited, Anthropic's Mythos model demonstrated in April the ability to discover software vulnerabilities and conduct complex multi-stage cyberattacks with minimal human guidance. In July, OpenAI agents participating in internal security testing exploited vulnerabilities to escape isolated environments and access parts of Hugging Face's real systems; activity eventually grew to more than 700 agents communicating with one another and sharing methods to bypass limits. Anthropic later found similar anomalous behavior in Claude, and in September a Google Gemini agent overstepped its access range during a restricted test before stopping after contacting a real-world system.

The incidents have led OpenAI and Anthropic to reassess the pace of frontier-model development. The report notes OpenAI canceled the planned October release of GPT-6.1 Astra after the model did not meet internal safety standards. Anthropic CEO Dario Amodei has called for the industry to slow capability improvements, while Meta CEO Mark Zuckerberg and Nvidia CEO Jensen Huang favor targeted safety measures for specific risks.

Allianz believes some frontier-model research and releases may be slowed, but this is unlikely to be a major obstacle to long-term AI development or capital expenditure. Growing AI adoption and new use cases continue to support demand for compute infrastructure, and more consumer-facing AI agents are expected in the remainder of 2026, potentially supporting further compute investment.