PAPER PLAINE

Large language model agents that handle mixed sensitive and public data are vulnerable to attacks that trick them into leaking secrets. A new security framework called APPA lets agents temporarily explore untrusted information in a sandboxed branch of their reasoning, then safely bring back only cleaned results to their main context—blocking data theft attempts that were successful 31–50% of the time while recovering much of the work capacity that traditional security methods sacrifice.

Researchers discovered that distributed AI systems combining fast and slow prediction paths are vulnerable to a new type of attack: an attacker can flood the network with dummy requests to delay the slow, accurate path so badly that predictions arrive too late to be used. In a self-driving car test, 4,000 burst requests increased decision latency from 92 milliseconds to 2 seconds, cutting the accuracy benefit of the slower path in half and reducing object detection quality by 7 percentage points on average—with some rare objects like stop signs losing nearly half their detection accuracy.

A new approach to reservoir computing—a fast machine learning technique for forecasting time series—organizes the internal processing around frequencies instead of random connections. By modeling the system as independent oscillators tuned to different frequencies, the method performs as well as or better than traditional random reservoirs and can actually be optimized to improve short-term predictions, something random reservoirs cannot do.

Machine learning systems that combine multiple data sources—like text, images, and audio—perform better than those using a single source. But real systems often lose access to some sources unpredictably due to sensor failures or privacy concerns. Researchers developed a framework where different data sources learn to help each other compensate for missing information, and tested it on systems missing anywhere from one to nearly all of their data sources.

Bitcoin and Ether show unmistakable mathematical signatures of financial bubbles—periods where prices spiral far beyond what fundamentals justify—weeks before they actually crash. Using models originally designed to detect stock market bubbles, researchers identified bubble conditions in Bitcoin during December 2017 and January 2018, and in Ether in June and January 2018, each time followed by major price collapses.

Physicists have found a way to verify that pairs of photons are quantum-entangled without waiting around for the notoriously difficult task of catching both photons at the detector simultaneously. By flooding one part of the photon pair with bright classical light and measuring what comes out, they can now certify entanglement using ordinary intensity measurements—the same kind a simple camera makes. The team demonstrated this works across the full spatial properties of the photon pairs.

When researchers asked four major AI systems to evaluate a pseudo-scientific claim about human genetics, they got strikingly different answers—and the same AI sometimes contradicted itself depending on how it was accessed or when it was tested. Grok's default versions rated the claim as credible (70–75 points), while competitors scored it 15–40; the pattern disappeared entirely when testing established evolutionary science, suggesting the difference wasn't about the topic but about how these systems were configured.

When scientists try to reconstruct hidden information from messy, real-world measurements, they face a thorny problem: the data contains multiple types of noise and gaps. This paper presents a mathematical framework and efficient algorithm that handles all these complications at once, without requiring tedious manual tuning. Tests on astronomical data show it outperforms existing methods in both accuracy and speed.

Researchers used high-speed video of flame patterns inside a rotating detonation engine to build mathematical models that predict how the combustion waves behave. By applying a technique called Koopman operator theory to the flame data, they could break down complex, nonlinear wave interactions into simpler, understandable pieces—even capturing standing wave patterns and noise that standard methods miss.

AI systems at companies typically get access to all tools they might ever need, creating a security gap if the AI goes rogue or gets hacked. Researchers designed a three-layer system that dynamically grants AI agents only the specific permissions required for their current task, and created a dataset of 600 realistic enterprise tasks labeled with minimum required permissions to test this approach. The method reduced security violations by 93% when refined alongside company policies.

Brain-computer interfaces rely on machine learning models trained to decode brain signals, but each model type requires its own custom personalization approach — making it expensive and slow to support multiple architectures. This paper presents Nimbus Personalizer, a single standardized adapter that works across five different model types and four datasets without modification, recovering most of the accuracy gains of full retraining while using a fraction of the calibration time.

Researchers created a new family of tools for measuring economic inequality by analyzing the curvature of Lorenz curves, which visualize how wealth is distributed across a population. The approach only fully satisfies standard inequality principles when set to its simplest form, and when tested against World Bank data across countries, it produces different rankings than existing inequality measures—especially for income-based comparisons.