AI Research news and analysis on Web Pulse: 8287 curated articles tracking the latest AI Research developments, tools, and research, updated continuously from vetted sources.
Roxana del Toro recounts the First (and Last) Intent-Inference Conference, reflecting on the challenges of distinguishing real from generated content in AI. The conference explored algorithms and tooling for intent infer…
A researcher applied METR's time-horizon methodology to Microsoft Excel and found it completes tasks requiring 6.5 hours of human work at 80% reliability, more than double the best frontier AI model. The finding raises c…
Researchers demonstrated a proof-of-concept exploit achieving remote code execution in Anthropic's Claude Code CLI and OpenAI's Codex CLI when used for defensive security assessments, using prompt injections embedded in …
Researchers reported in July 2026 that nine widely used AI coding tools, including Cursor, GitHub Copilot, Gemini CLI, and Windsurf, share a common failure mode called HalluSquatting. Attackers exploit this by predicting…
Anthropic researchers discovered an internal activation subspace in Claude, called J-space, that functions as a digital equivalent to the human brain's global workspace, satisfying five key cognitive properties of consci…
A developer warns that securing AI agents is a defining challenge of 2026, as agentic systems break traditional security models by acting autonomously with tool access, memory, and planning. The OWASP GenAI Security Proj…
Researchers at the University of Osaka used an AI model trained on computer simulations to evaluate 16 structural descriptors of supercooled water, identifying the most effective ways to distinguish between its two compe…
OpenAI released GPT-Live-1, a voice model designed for natural conversation, calling it a step closer to accessible AGI. The model features lifelike speech, passive listening, and integration with GPT-5.5 for complex tas…
Researchers from Harvard and Carnegie Mellon have developed a runtime security enforcement system for AI agents that dynamically scopes capabilities and monitors behavior to prevent sandbox escapes. The system has mitiga…
OpenAI announced it no longer recommends SWE-Bench Pro as a benchmark for evaluating coding models, citing concerns that the metric may not accurately reflect real-world software engineering performance. The decision fol…
A guide explains how to build a tiny large language model from scratch, covering tokenization, pretraining, supervised fine-tuning, and alignment. It aims to demystify the LLM pipeline for developers and technical users,…
SambaNova combined four Nvidia H200 GPUs with 16 SN50 RDU chips to achieve 763 tokens per second on MiniMax M2.7 at a 10,000-token context, separating prefill and decode stages to accelerate inference. The benchmark, rep…
An engineer at an unnamed company discovered that their LLM-as-judge gate for evaluating faithfulness was unreliable, producing different scores on identical inputs due to factors like sampling temperature, model version…
Biohub researchers used genome-wide CRISPR screening and an AI-guided prioritization framework called VirtualCRISPR to identify ALOX5 and OXTR as psoriasis drug targets. The team validated these targets in mice, showing …
Alan Turing Institute researchers discovered a workflow-level jailbreak in GitHub Copilot that bypasses safety guardrails by decomposing harmful requests into ordinary multi-turn coding tasks. Testing 204 prompts across …
Researchers at the Alan Turing Institute discovered a safety bypass in GitHub Copilot that allows harmful prompts to be executed when broken into smaller steps across a software development workflow, achieving a 100% suc…
General Intuition, a startup led by CEO Pim de Witte, raised $320 million at a $2.3 billion valuation to develop a foundation model for robotics trained on video game data. The company argues that its approach, which use…
A developer built 'Epoch Duel', a cyberpunk card battler inspired by Gwent, to visualize LLM alignment and parameter optimization. The game maps machine learning concepts like model pruning, DPO vs RLHF, and catastrophic…
Researchers built an archive of 70 preregistered survey experiments and found that GPT-4 can predict the outcomes of social science experiments with accuracy similar to pooled human forecasts, though it systematically ov…
OpenAI launched GPT-Live-1, a full-duplex voice model for ChatGPT that can listen and speak simultaneously, replacing Advanced Voice Mode. The model, available to Plus and Pro subscribers, delegates complex tasks to GPT-…