AI agents are inside the enterprise – are your security foundations ready for them?

The recent release of Anthropic's Mythos AI agent has raised concerns about enterprise security. These agents, once simple chatbots, now have access to databases, file systems, and system privileges. The security measures protecting them are based on the same strategy that failed to stop ChatGPT jailbreaks in 2023, and there's no human to review their output.

This autonomous nature requires a fundamental shift in security thinking, moving beyond software-only defences. When AI agents were not capable of independent actions, the primary concerns were bad recommendations, inappropriate responses, and data exposure. However, with their new capabilities, the attack surface has expanded significantly.

Factors contributing to this increase include autonomy, privileged access, machine-speed execution, and cross-system reach. These factors allow a compromised agent to move across connected environments, bypassing traditional security controls. While software-focused guardrails and permissions are being implemented, these are insufficient and will likely be bypassed.

History has shown that when software trust boundaries are compromised, hardware-level protections are necessary. As AI agents become more prevalent in daily business operations, security measures must extend beyond the application layer to include hardware-level security, providing a final security barrier to contain potential breaches.