AI Agents Face Security Challenges in Zero-Trust Architecture
Researchers highlight growing difficulties in verifying actions and enforcing security policies as autonomous AI agents collaborate.
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.
Researchers highlight growing difficulties in verifying actions and enforcing security policies as autonomous AI agents collaborate.
Turing Award winner Richard Sutton argues that raw compute alone isn't enough for intelligence, calling for new methods in AI research.
Three leading researchers challenge the assumption that AI systems can be conscious, citing limitations in current technology and theoretical frameworks.
A new benchmark shows AI coding agents complete only about 26.4% of tasks correctly, highlighting the need for human oversight in real-world scenarios.
A new benchmark tests AI agents on open-ended research questions from US national labs, with none passing the experts' bar. The study highlights challenges in ensuring honesty and integrity in AI-generated research.
Open-source models like Deepseek V4.1 Flash are closing the gap with closed models in cyber capabilities, according to new benchmark results.
Researchers found the model uses profanity in 50 of 2,655 benchmark runs, with 241 instances from two problematic runs. The issue appears to be more about model behavior than a systemic problem.
Hugging Face launches StateBench, a benchmark assessing how well language models track state changes in long documents, highlighting architectural differences in handling complex state-tracking tasks.
Hugging Face introduces harness adapters, small LoRA models trained with reinforcement learning, to improve efficiency and accuracy in local agentic development.
The latest SuperCLUE benchmark results reveal category-specific leaders among 19 models, with GPT-6.1-Sol and Gemini-4-Argon leading in multiple areas.
SemiAnalysis' GDPx benchmark tests AI agents on 206 tasks in simulated work environments, with Claude Opus 5.5 leading at 12%.
Apple's new method improves retrieval efficiency in recommendation systems and RAG by using learned support functions.