Photon Framework Boosts AI Vulnerability Detection

ORNL's Photon leverages Frontier's exascale speed to rapidly identify weaknesses in AI models.

Mar. 25, 2026 at 2:22am

Got story updates? Submit your updates here. ›

Researchers at Oak Ridge National Laboratory's Center for Artificial Intelligence Security Research (CAISER) have developed Photon, a framework that can quickly detect vulnerabilities in AI models by running multiple attack scenarios simultaneously on the Frontier exascale supercomputer. Photon adapts and evolves its tactics in real-time to uncover both obvious and subtle weaknesses, providing a comprehensive understanding of model performance under adversarial conditions.

Why it matters

As AI models are increasingly deployed across critical domains like energy, healthcare, finance, and national security, ensuring their reliability, robustness, and safety has never been more vital. Photon offers a way to proactively identify and address AI vulnerabilities, safeguarding the integrity and performance of mission-critical systems.

The details

Photon is built on ORNL's existing DeepHyper technology, which was originally developed for training large neural networks. By inverting its purpose, Photon is now trained to detect nefarious activity, allowing it to rapidly identify the most efficient attack parameters against AI models. Photon begins by applying known attacks from published literature, then refines and exploits the discovered vulnerabilities while also exploring the model further to uncover new weaknesses. This cycle continues until no further degradations in the model's performance are observed.

  • Photon was developed by researchers at ORNL's Center for Artificial Intelligence Security Research (CAISER).
  • Photon is designed to run on ORNL's Frontier exascale supercomputer, which was recently installed.

The players

Photon

A groundbreaking framework developed by ORNL researchers to rapidly detect vulnerabilities in AI models at exascale.

CAISER

The Center for Artificial Intelligence Security Research at Oak Ridge National Laboratory, which is focused on identifying and characterizing vulnerabilities in AI models.

Frontier

ORNL's exascale supercomputer, which enables Photon to run multiple AI vulnerability scenarios simultaneously at high speed.

DeepHyper

ORNL's existing technology that Photon is built upon, originally developed for training large neural networks.

Edmon Begoli

The director of CAISER at ORNL, who oversees the development of Photon.

Got photos? Submit your photos here. ›

What they’re saying

“It might sound devious, but it's worked very well. Photon accelerates the design and development process and reuses the most effective methods for exploring and exploiting vulnerabilities.”

— Edmon Begoli, Director of CAISER, ORNL

“When we're talking about running something at this scale, it becomes difficult to use as much of the available compute power as possible. Since you are running at such a large scale, eliminating resource downtime is not trivial. There is still downtime when resources are waiting for what to do next but maintaining 95 percent utilization is very high.”

— Jack Hutchins, Robust AI Engineer, ORNL

“Since our goal is to find highly effective jailbreaks, finding the parameters that have the most effect quickly speeds up our search for effective jailbreaks.”

— Jack Hutchins, Robust AI Engineer, ORNL

“Photon represents a paradigm shift in how we approach AI security. By running coordinated, high-scale experiments, we can uncover hidden vulnerabilities far more efficiently than ever before. This technology ensures our AI advancements can continue bringing much-needed innovation to a wide variety of industries without also introducing safety or security risks.”

— Edmon Begoli, Director of CAISER, ORNL

What’s next

Photon's development and testing on the Frontier supercomputer will continue, with the goal of making the framework available to AI model developers and researchers to proactively identify and address vulnerabilities in their systems.

The takeaway

Photon's ability to rapidly and comprehensively detect vulnerabilities in AI models at exascale represents a significant advancement in ensuring the reliability, robustness, and safety of critical AI-powered systems across industries. By uncovering both obvious and subtle weaknesses, Photon helps safeguard the integrity of AI-driven innovation.