AI Has A Trust Problem. The Answer Is In The Chip

The release of Claude's Mythos model served as a stark warning: AI cannot be ignored or fully trusted—at least not yet. In April, Anthropic gave approximately 50 companies that build or maintain critical software access to Mythos through Project Glasswing. Within weeks, the model discovered a 27-year-old OpenBSD vulnerability and a bug in FFmpeg that five million automated tests had missed.

While this capability benefits defenders, it also empowers attackers. Just a month later, Google's Threat Intelligence Group confirmed the first known zero-day exploit built with AI assistance. These cases reveal a dangerous shift: the time to find security flaws and exploit zero-day vulnerabilities has collapsed to machine speed.

The Speed Of Exploits

The Zero Day Clock measures the time-to-exploit window from vulnerability disclosure to active exploitation. In 2018, this median time stood at 771 days. By 2025, the window had shrunk so dramatically that most vulnerabilities were weaponized before they were even disclosed.

Models like Mythos have further shifted the risk balance in attackers' favor by exposing more vulnerabilities than can be timely patched. In the datacenter era, both discovering and exploiting vulnerabilities required manual effort. Cloud and SaaS automated only the exploitation phase. Now, AI has automated both steps, effectively closing the response window.

Agents Inside The Gates

The second critical factor involves who holds insider access. Early enterprise computing granted privileged access to a small set of employees. Cloud and SaaS extended this to vendor engineers with standing credentials. Today, AI agents hold access to systems and data on behalf of those users.

Many assume AI agents have limited access to sensitive data, but the opposite is typically true. This manifests through three challenges: non-human identity sprawl, data-in-use exposure, and third-party model dependency. In customer environments, seemingly benign coding agents can reach into users' local environments, read files and environment data, and pass that content back to the model.

AI-First Security Approach

Traditional security approaches no longer suffice against AI-accelerated threats. Perimeter and identity tools were built for human pace, not autonomous agents. Encryption protects data at rest and in motion, but not when an agent reads it. Zero trust assumes compute layer integrity, and AI governance documents intent without enforcement.

One mitigation strategy is favoring protection over detection. Confidential computing can help keep data protected even with zero-day vulnerabilities in the hypervisor or operating system. This approach runs workloads in hardware-isolated, encrypted enclaves built into modern CPUs and GPUs, where applications and data remain sealed from everyone else on the machine, including administrators and cloud operators.

Defending Against AI Threats

Organizations must take specific steps to defend against AI-accelerated threats: map exploit-lifecycle exposure, treat every agent as a non-human identity with least privilege, and recognize that patching alone is insufficient. They should use technologies that protect data in use, demand proof rather than policy from AI supervision tools, and monitor frontier AI security tools for dual-use risk indicators.

The deciding factor isn't whether an organization adopts AI, but whether its trust model can be verified rather than assumed. As AI continues to evolve, hardware-based security solutions will become increasingly critical to maintaining trust in an era of machine-speed threats.