Anthropic's Mythos Model Identifies Vulnerabilities in Internet Security Algorithms
Anthropic's AI model found weaknesses in cryptographic algorithms, including a reduced version of AES, with no immediate impact on current systems.
387 articles
In-depth coverage of new AI research — papers, benchmarks, and breakthroughs from leading labs and academia, summarized for fast reading and grounded in the methods that matter.
Anthropic's AI model found weaknesses in cryptographic algorithms, including a reduced version of AES, with no immediate impact on current systems.
Google's new study of 15 million AI interactions finds AI use remains shallow and collaborative, with less than 25% of work tasks significantly impacted.
Rohde & Schwarz highlights the limitations of static threat libraries against mode-agile emitters, which now dominate modern RF environments.
AWS announced task-aware knowledge compression (TAKC) to enhance enterprise AI workflows by pre-compressing knowledge bases into task-specific representations, reducing token count by 8x to 64x.
METR's new metric calculates when AI agents exceed human cost-effectiveness, estimating $2,500 per one-percent speedup in human effort.
NVIDIA outlines the growing role of simulation in training physical AI systems, with a focus on robotics and control policies.
Anthropic's Claude Opus 5 scored 30.2 percent on the ARC-AGI-3 benchmark, nearly four times the previous record of 7.8 percent set by OpenAI's GPT-5.6 Sol (Max).
Kimi K3 scored 32.2% on ExploitBench, while leading U.S. models reached 76.2% in cyber exploit development, per a joint evaluation by the British AI Security Institute and the U.S. Center for AI Standards and Innovation.
Anthropic announced a $200 million fund to support research on AI's economic impacts, focusing on worker adaptation and policy interventions.
An AI system helped Pakistani judges clear 1,848 more cases per year per district, yielding a $38.50 return per dollar invested, according to a study.
Researchers suggest new AI evaluation metric to measure alignment with user intent, citing 80% of users struggle with current systems.
Amazon Nova 2 uses self-distilled reasoning to improve performance on math and coding tasks, recovering 70% of base model performance after SFT training.