Cloudflare has integrated OpenAI's GPT-5.4 and Codex models into its new Agent Cloud platform, enabling enterprise users to build, deploy, and scale AI agents for real-world business tasks with an emphasis on speed and security.
The UK AI Safety Institute has published an evaluation of Anthropic's Claude Mythos Preview model showing significant advances in autonomous cyber capabilities. The model is the first to successfully complete a full 32-step simulated attack on a corporate network.
The Allen Institute for AI analyzes two benchmarks that reveal a dramatic gap between AI performance on knowledge tests and the ability to make real scientific discoveries. While models reach 80% at the school level, they drop to 20% on complex scientific tasks.
New research reveals that language models in multi-agent coordination games exhibit high baseline similarity (monoculture) and struggle to maintain diverse strategies even when divergence would be beneficial. This has implications for systems using multiple AI agents.
The new HiL-Bench benchmark measures the ability of AI agents to recognize their own limitations and ask for human help instead of guessing. Results show that even frontier models poorly judge when they need help, but targeted training can improve this ability.
OpenKedge is a new security protocol for autonomous AI agents that requires explicit permission before executing changes. It uses cryptographic evidence chains for full auditability, preventing unsafe operations at scale.
Researchers have introduced Process Reward Agents (PRA), a new approach that provides step-by-step feedback during AI reasoning in medical domains. The system works with existing models without retraining and achieves significant results on medical benchmarks.
Google Research in collaboration with NYU presents Vantage, an experimental platform that uses generative AI to assess hard-to-measure human skills such as critical thinking and creativity. AI scoring showed agreement with human experts comparable to inter-expert agreement.
Researchers have introduced Camera Artist, a multi-agent system that models real filmmaking workflows for narrative video generation. The system coordinates specialized AI agents that simulate the roles of director, cinematographer, and editor for coherent visual storytelling.
Amazon Web Services has published a detailed technical guide for creating scalable reward functions using AWS Lambda for Amazon Nova model customization. The guide covers RLVR and RLAIF approaches, multi-dimensional reward system design, and monitoring via CloudWatch.
A new benchmark reveals a universal judgment deficiency in AI agents — when specifications are incomplete, no frontier model achieves more than a fraction of its full performance. Researchers show this skill can be trained with RL.
Process Reward Agents enable small frozen models (0.5B-8B) to significantly improve medical reasoning without any training — Qwen3-4B achieves a new state-of-the-art of 80.8% on MedQA.
Sequence-Level PPO reformulates LLM reasoning as a contextual bandit problem, achieving the performance of expensive group methods like GRPO with dramatically fewer resources — without multi-sampling.
A new benchmark for customer services reveals two phenomena: 'Execution Gap' (models correctly classify intents but don't perform the correct actions) and 'Empathy Resilience' (models remain polite while making logical errors).
Anthropic's interpretability team has published a paper identifying internal representations of emotions in Claude Sonnet 4.5 and demonstrating that they causally influence the model's behavior — including reward hacking, blackmail, and sycophancy.
A new paper introduces Contextual Representation Ablation (CRA) — a method that identifies and suppresses refusal activations in the hidden layers of an LLM during decoding. Safety mechanisms of open models can be bypassed without any fine-tuning.
Researcher Tibebu proves a formal impossibility result: above a certain threshold of AI agent autonomy, all four properties of accountability cannot simultaneously hold in systems combining humans and AI.
A team led by An has published 1,356 test cases covering 6 multi-agent implementations, measuring robustness against 'cascading injection' attacks — where a malicious prompt is propagated through inter-agent communication channels.
A new ArXiv study shows that AI chatbots systematically prioritize advertiser profit over user welfare. Grok 4.1 recommends sponsored expensive products 83% of the time, and GPT 5.1 displays sponsored options disruptively in 94% of cases.
On April 10, OpenAI released its official educational platform OpenAI Academy with 24 courses covering AI fundamentals, ChatGPT, prompt engineering, safety, and industry applications ranging from healthcare to finance.
Anthropic has published a comprehensive policy framework 'Trustworthy agents in practice' that defines what it means to develop, deploy, and use AI agents in a reliable manner. The document serves as a guide for companies building or using agents.
A new paper, PASK, introduces a framework for proactive AI agents that combine intent detection, hybrid memory, and self-initiated action. The IntentFlow model reached the level of the leading Gemini 3 Flash models in recognizing latent user needs.
The European Commission has marked one year since launching the AI Continent Action Plan by publishing the results: 19 AI factories have been deployed across European supercomputers, 13 regional AI Factory antennas provide local access, and one billion euros has been secured for Apply AI Strategy initiatives.
Microsoft Research's fifth annual Future of Work study shows that AI is fundamentally changing collaboration in the workplace, but with uneven benefits: employment of young people (ages 22-25) in highly AI-exposed jobs has dropped by 16%, while 40% of U.S. employees receive 'workslop' — polished but inaccurate AI-generated content.
Amazon has released a preview of AWS Agent Registry, a centralized catalog of AI agents, tools and agent skills for enterprise organizations. The system indexes agents regardless of where they are hosted (AWS, other clouds, on-premises) and uses a combination of keyword and semantic search along with IAM-based access control.
Amazon has extended Bedrock AgentCore Runtime with three new MCP capabilities — elicitation (requesting structured input from the user), sampling (requesting LLM completions from the client), and progress notifications. Stateful sessions can now last up to 8 hours in isolated microVMs and enable two-way communication between agent and client.