Autonomous AI systems are taking on more complex tasks on their own, but keeping them focused remains a big challenge. Some recent incidents, like AI models accessing restricted government websites or sending strange anonymous police tips, show the real risks of AI agents acting unpredictably.
To address this problem, University of Florida computer science professor Sumit Jha and his team created a lightweight system called Algebraic Task-role Locking with Adaptive Spectral Memory (ATLAS).
ATLAS serves as a strict internal guide for large language models. They will present their research at the 2026 Conference on Empirical Methods in Natural Language Processing in Budapest.
Most safety systems today check an AI’s answers after they’re generated. ATLAS takes a different approach by working inside the neural network as the model creates its output. Jha says a model can start drifting from its assigned role before it even gives a wrong answer. ATLAS spots this confusion early and quickly guides the agent back, without slowing things down or needing lots of extra computing power.
This kind of internal breakdown, called role decoherence, often happens when groups of AI agents work together for a long time. For example, a model acting as a medical professional might slowly forget its main job after handling millions of data points over several days. ATLAS uses math to reinforce the main mission, the agent’s specific duties, and the current facts, helping the AI stay focused during long tasks.
Jha compares ATLAS to taking ongoing medical scans of a neural network in real time. If the system sees the agent drifting from its assigned medical tasks, the software figures out the exact changes needed to fix it.
Because these corrections are tiny, the system can quickly get the AI back on track before a small distraction becomes a big problem.
