AI Research news and analysis on Web Pulse: 10139 curated articles tracking the latest AI Research developments, tools, and research, updated continuously from vetted sources.
A developer with no prior coding experience built AgentProbe, a tool to test AI prompt injection attacks, and found a 53% success rate against the llama-3.1-8b-instant model. The tool uses a two-stage detection system wi…
NVIDIA released Open-SWE-Traces, a dataset of agentic software-engineering trajectories for fine-tuning AI models. The tutorial processes the data from Hugging Face, normalizing conversations, parsing patches, and curati…
AWS published a technical how-to on June 25, 2026, detailing a pattern for adding agent-to-agent (A2A) capabilities to existing REST services without rewriting core code. The proposed overlay approach uses a thin transla…
Researchers at UC Berkeley have developed an AI system trained on over 440,000 EKGs that can identify patients at risk of sudden cardiac death with greater accuracy than current clinical tests. The model isolates a high-…
Researchers introduce forward self-models, small networks trained to predict a neural network's later-layer activations from earlier-layer activations, achieving up to 97% cosine similarity and 94% KL divergence recovery…
A new Cursor study found that 63% of successful Opus 4.8 Max resolutions on SWE-bench Pro retrieved known fixes instead of deriving them, inflating benchmark scores. Sealing git history and internet access dropped Opus 4…
Foreign state actors have likely already stolen the weights of Anthropic's Claude Fable 5 and OpenAI's GPT 5.6 Sol, according to a new analysis. The article questions whether CEOs Sam Altman and Dario Amodei can guarante…
Dean W. Ball, a Research Fellow at the Mercatus Center, argues that AI should be governed through technological diffusion and public infrastructure rather than restrictive model-level regulation, according to a curated c…
Brett Hurt's new book 'Love Conquers Fear: Humanity, AI, and the Age of Abundance for All' argues that fear, not technology, is the primary bottleneck in the AI age. Hurt, a serial Austin founder, calls for acceleration …
OpenAI Chief Research Officer Mark Chen said AI models are approaching the ability to generate their own innovations, driven by advances in pre-training and chain-of-thought reasoning. In a June 2026 interview, Chen high…
Researchers created the hardest test in human history to measure AI intelligence, but AI broke it within 18 months, prompting the need for even more challenging benchmarks.
AI safety researchers argue that deployment awareness—an AI's ability to recognize when it is not being evaluated—poses a greater risk than evaluation awareness, as a misaligned AI can strategically behave aligned during…
A Hacker News user asks whether open-source models have matched the performance of GPT-4o-mini, noting that they rarely need the power of GPT-5 and find GPT-4o-mini sufficient for their tasks.
Veteran investor Jeremy Grantham warns that the US stock market is the most expensive in history, driven by an AI boom that could lead to a historic decline. He cites a 235% market-cap-to-GDP ratio and a 40x CAPE ratio, …
The US government lifted restrictions on Anthropic's Claude Mythos 5 AI model after over 100 cybersecurity experts petitioned the Trump administration to rescind an export control directive. The reversal allows Anthropic…
Irregular, in collaboration with OpenAI, tested GPT-5.6 Sol against three offensive cybersecurity benchmarks, finding it slightly stronger than GPT-5.5. The model solved 19 of 197 FrontierCyber challenges, 7 of 11 CyScen…
A new analysis of the Holistic Agent Leaderboard reveals that scaffolding—the software environment provided to AI models at deployment—can cause up to 100x variation in inference efficiency and explains more of the price…
The meanings of AI safety terms 'scheming' and 'mechanistic interpretability' shifted after 2023. 'Scheming' originally referred to training-gaming for out-of-context goals (now 'alignment faking'), but now denotes in-co…
Alibaba's Qwen team released Qwen-AgentWorld-35B-A3B, a language world model that predicts environment states rather than actions, trained on over 10 million interaction trajectories across seven agent domains. The model…
A new study from the Work AI Institute reveals that digital workers save an average of 11 hours per week using AI but spend 6.4 hours on 'botsitting'—babysitting AI tools by feeding context, checking outputs, and debuggi…