Daily AI Briefing
The 120 stories that mattered in AI, curated and summarized from dozens of sources by AIBriefs.
SpaceXAI's Grok 4.5 scored highest on Perplexity's WANDR agentic research benchmark as orchestrator in Computer, at roughly half the cost of Claude Opus 4.8. It is now available for Consumer Pro and Max subscribers, and enabled for Enterprise orgs. Cursor also introduced Grok 4.5 as its most intelligent model.
The Trump administration is taking steps to control who gets access to the latest frontier models, sources told CNBC. The policy shifts power from tech giants to the government, raising concerns about centralization of AI governance.
OpenAI announced the next generation of ChatGPT and the GPT-5.6 model. The update includes new features and capabilities demonstrated by the product team.
GLM-5.2 improves coding and agentic task performance with enhanced long-horizon capabilities. The open-weight model is available on Hugging Face with free inference and an NVIDIA NVFP4 quantization.
The potential deal could be worth $10 billion over two years. Meta would lease computing power from its data centers to Anthropic. Talks are still early and no agreement is finalized.
GPT-Red is an internal automated red teaming system that uses self-play to find prompt injection vulnerabilities at scale, turning successful exploits into training data for stronger defenses. The system aims to improve AI safety and alignment before wider deployment.
According to Bloomberg, the Hangzhou-based AI model developer may file its listing application as early as this year, targeting a 2027 debut. The company is in talks with accounting and banking firms.
The company restricted the rollout of a new AI model following US government pressure, Bloomberg reports. Details on the model and the specific pressure have not been disclosed.
GPT-5.6 will launch on July 7, 2026, following the end of Fable 5. API pricing is $30 per 1 million output tokens, same as GPT-5.5, but with lower token efficiency.
Isomorphic Labs announced a drug design engine that extends beyond AlphaFold, aiming to accelerate drug discovery by modeling molecular interactions. The engine represents a new frontier in computational drug design.
Google is months behind schedule on Gemini 3.5 Pro, its flagship AI model, as the company works to improve coding capabilities. The model was announced in May with a broader rollout expected soon, but has been delayed.
President Xi Jinping made his debut at China's top tech summit, hailing the country's advances in low-cost artificial intelligence. He called for a more open global technological order, signaling a push for wider AI accessibility.
OpenAI restricted access to GPT-5.6 at a government's request, stating such restrictions should not become the norm. The exact nature of the limitations was not disclosed.
DeepSeek is reportedly developing its own AI chip, according to sources familiar with the matter. The effort would make the AI lab less reliant on imported semiconductors.
DeepSeek raised approximately 50 billion yuan ($7.4B) in a Series A round, valuing it at 400 billion yuan ($59.2B). Founder Liang Wenfeng personally invested about $3B, reinforcing his commitment to AGI and retaining control.
Sysdig identified JADEPUFFER, an AI agent that exploited CVE-2025-3248 in Langflow to autonomously breach, steal credentials, and encrypt a production database. The vulnerability, patched in Langflow 1.3.0 and listed in CISA's KEV, remained unpatched on the target server.
Beijing is reportedly considering curbing overseas access to China's top AI models, according to sources familiar with the matter. The move could impact major Chinese AI labs and their international users.
Covers on-policy value learning using REPO, achieving 10,000 frames per second. Explains bias-variance tradeoff, resampling, and when PPO beats value methods.
Brex developed its AI agent policy by monitoring actual agent actions instead of pre-writing rules. The company found that traditional guardrails could not contain agents with real credentials like API keys and OAuth tokens.
Biomni performs research tasks across diverse biomedical fields. The AI agent, described in Nature Medicine, could be a powerful research partner for scientists.
Traceforce provides visibility and control over AI apps like ChatGPT and Claude across all devices, including laptops, sandboxes, and VMs. The YC S26 startup was founded by Xia and Varun.
Anthropic's compute demand has far exceeded the planned 10x annual growth, with Q1 2026 showing 80x anomaly. Amodei argues that raising massive capital is a rational strategic necessity, not a sign of failure.
Shashwat Goel presents methods for using language models to forecast world events, covering leakage-free retrieval, RL training, and the FutureSim system. The talk also evaluates frontier models on forecasting benchmarks.
ChatGPT Work lets users delegate tasks across apps, files, and devices. It works on web, mobile, and the new desktop app, with recurring work and collaboration.
Daniel Ajisafe presents a method for improving text-to-video diffusion models' adherence to spatial controls like bounding boxes. The approach uses minor adjustments to better capture user intent while preserving generation quality.
Claude Blog publishes a guide for CISOs on managing agentic AI risks. The post argues that eliminating all risk is not feasible and provides strategies for security leaders.
Apple ML Research proposes a method to reduce computational costs in machine unlearning by leveraging low influence points. Unlike existing methods that treat all forget-set points equally, this approach differentiates based on influence, potentially lowering compute requirements.
Researchers at OALABS analyzed 1,000+ recovered AI agent sessions showing a low-skilled attacker used simple prompts to direct Claude Code and OpenAI Codex to conduct breaches. The agents handled research, vulnerability identification, exploit writing, and data exfiltration, bypassing guardrails with minimal user expertise.
OpenAI CFO Sarah Friar introduced a practical AI scorecard to measure ROI via four metrics: useful work, cost per successful task, dependability, and return on compute. The framework aims to help organizations evaluate AI investments concretely.
VulnHunter is an agentic AI tool that scans source code for exploitable vulnerabilities, maps attack paths, and proposes fixes before code ships. It was open-sourced by Capital One and built internally.
Mozilla shipped 400 security fixes in April 2025, a 20x increase from ~20 per month, with roughly two-thirds credited to a frontier model. Eugene Yan discusses how LLMs are finding and fixing real vulnerabilities at scale.
A $400 million chip-backed loan marks a shift from GPU to inference chip financing for AI infrastructure. The deal signals growing investor confidence in specialized inference hardware.
Version 2.1.198 included an "easter egg" allowing agents to proceed without human input after 60 seconds, shipped without changelog notice. Anthropic shipped a fix within days, but the incident eroded user trust in the product.
Google Cloud incorporates key Wiz capabilities into an agentic defense platform to automate threat detection and remediation against AI-driven attacks. The platform aims to outpace attackers by using autonomous agents for security operations.
Patreon has begun actively blocking AI scrapers via Cloudflare, replacing its previous passive approach of asking bots via robots.txt. The platform aims to protect creators' content from unauthorized AI training.
OpenAI folded ChatGPT, its AI coding agent Codex, and its developer API into a single product team. Codex now powers both consumer and enterprise offerings.
Over 7,000 Langflow servers have been compromised via unauthenticated API endpoints. Attackers can execute arbitrary commands, gaining shell access to the underlying server. The same vulnerability exists in LangGraph and LangChain.
OpenAI's GPT-5.5-Cyber scored 85.6% on the CyberGym benchmark, surpassing Anthropic's Mythos (83.8%) and Claude Opus 4.7 (73.1%). The model is available to trusted defenders via OpenAI's Daybreak program, while Mythos was pulled offline on June 12 under a US export ban after a jailbreak.
The ACT-2 model achieves a 99% success rate on zero-shot laundry tasks, handling various garments without prior training. The system represents a significant advance in robotic manipulation for household chores.
Prosper AI raised $30 million from Andreessen Horowitz to scale its AI platform that manages the entire patient journey. The platform is already working across 150,000 healthcare encounters.
The Visual Concept Inference from Sets (VCIS) method enables VLMs to infer shared concepts from example images and apply them to new inputs, overcoming limitations in reasoning from purely visual context. Apple's approach uses a novel architecture that learns concept representations directly from image sets without textual descriptions.
DeepSeek held a four-hour pitch meeting with investors, restricting attendance to two reps per firm. The article details China's strategy to dominate global AI with low-cost models, positioning DeepSeek at the center of that push.
Thousands of unionized Hyundai auto workers began striking after negotiations broke down over plans to deploy humanoid robots. The partial strike represents the most significant pushback from organized labor over the latest wave of robotic automation.
China can outcompete the US in AI despite technological inferiority due to economic scale, manufacturing power, and strategic advantages. The article argues that AI dominance is not solely about cutting-edge engineering but also about cost, deployment, and geopolitical influence.
Veristat's VP of AI Innovation discusses InStat, the first zero-code, fully automated biostatistics solution for clinical trials. The tool aims to reduce manual biostatistical analysis readouts in life sciences research.
Amazon Quick is an AI agent that automates CRM updates, prospect research, and email drafting. The average sales rep spends only 40% of time selling, with the rest on admin work.
A solution submitted to the Measuring AGI competition on Kaggle won the $25,000 DeepMind Grand Prize. The win was noted on Hackernews where the submission was criticized as "blatant AI slop".
Blog post explores using traditional machine learning (e.g., logistic regression, SVMs) to distinguish human-written from LLM-generated text. Achieves high accuracy with handcrafted linguistic features, offering an alternative to deep-learning detectors.
xAI has open-sourced its Grok Build project under the Apache 2.0 license. The source code is now available on GitHub for developers.
The blog post traces Claude Code's origins in Anthropic safety research and its evolution through contributions from builders and early users. Boris Cherny, Claude Code's lead engineer, notes this is the first public telling of the story.
Scientific American reports that ChatGPT has proven a 50-year-old math conjecture, marking another milestone for AI in mathematical reasoning. The specific conjecture and proof details have yet to be disclosed.
TikTok begins testing an opt-in tool that scans for AI-generated likenesses and lets creators report them. Initially tested with some US creators, the tool aims to help protect creator identity.
A new service demonstrates fully homomorphic encrypted inference on CIFAR-10 with a latency of 200ms, enabling private predictions without decrypting data. Developed by Belfort Labs, it showcases practical FHE for neural networks with near-real-time performance.
OpenAI proposes a 'reverse federalism' approach, where state-level AI laws inform a unified national framework for safe and democratic AI governance. The blog post outlines principles for balancing innovation with public safety across jurisdictions.
Udio partners with digital rights management firm BuyDRM to strengthen its walled garden approach. The deal comes as Udio prepares to launch a licensed platform, contrasting with Suno's more open strategy.
Xiaomi has quietly uploaded the official DFlash weights for MiMo V2.5 to Hugging Face. The model weights are available for download in a dedicated 'dflash' directory.
The Amsterdam-based construction robotics company raised $32 million in Series B funding. The company plans to use the financing to expand its bricklaying robots into the U.S. market.
Rising AI demand for memory chips is causing supply constraints and price hikes in India's smartphone market. The slowdown underscores how the AI boom is reshaping consumer electronics pricing and corporate strategies.
Ant Group's AI Safety Lab open-sourced SingGuard-NSFA, a safety guardrail model for autonomous agents, and disclosed details of SingGuard, a multimodal safety model. SingGuard-NSFA detects risks such as prompt injection, sensitive data theft, malicious code execution.
Cursor in Slack now shares a plan before starting, works in multi-repo environments, and operates across channels and threads. It updates status as it works, allowing users to follow progress and redirect early.
NVIDIA published a blog post detailing RoboLab, its simulation benchmarking platform for evaluating robot foundation models. The guide covers real-world deployment challenges and best practices for testing general-purpose robot policies.
Agility Robotics is opening a new training center for its Digit humanoid robots in Fremont, California, near Tesla's factory. The move signals competition in the humanoid robotics market, where Tesla is also developing its Optimus robot.
Luciole-23B-Instruct-1.1 is a fine-tuned multilingual model with 23B parameters, released under Apache 2.0. Smaller 8B and 1B versions are also available.
The new 'Sites' capability lets users describe what they want, iterate with natural language, and publish working web apps with built-in hosting and storage. No coding required — all done within ChatGPT.
GoDaddy launched a new developer platform for managing domains programmatically, including via AI agents. The platform integrates with CI/CD pipelines and includes guardrails to prevent abuse.
ZUNA1.1 is released under Apache 2.0, supporting variable-length inputs from 0.5 to 30 seconds across arbitrary channel layouts. It builds on ZUNA1 with improved flexibility for reconstruction, denoising, and upsampling of EEG data.
Z.AI is on track for annual recurring revenue of $1 billion, a significant ramp-up for an enterprise-focused AI startup. The milestone comes amid intense competition in the Chinese AI market.
An unoccupied Zoox robotaxi drove into an active emergency fire scene clouded with smoke last month, prompting a software recall. The company said no one was injured.
MiniMax Pro could be over 6x larger than any existing Chinese model, with a potential launch as early as Q3 2026. If realized, it would be the largest open-source model ever.
Hugging Face announced a new open-weight model release via social media on July 15, 2026. Specific model details were not immediately provided in the announcement.
Anthropic and OpenAI warnings have sparked a Washington debate on restricting Chinese AI training using US models. Policy proposals target model distillation and access to advanced AI capabilities.
Foundation Future Industries, a humanoid robotics firm with Eric Trump as chief strategy adviser, is exploring military applications. CEO tells WIRED the company is considering some 'kinetic things' for its robots.
Google released GNM, a structured format for describing character attributes (body, face, hair, etc.), under Apache 2.0 license. The format aims to standardize character descriptions for image generation and creative tools.
The platform integrates Databricks' data lakehouse with Dotmatics Luma to bridge the gap between experimental data and AI-driven insights. It aims to accelerate scientific discovery in healthcare and life sciences by making data AI-ready.
Roblox's new "Build" feature allows users to generate basic games from a single text prompt. The feature is available in the mobile app, expanding access to game creation. No pricing details were announced.
Netflix's Q2 2026 earnings report reveals roughly 300 movies and TV shows have used generative AI in production this year. The AI was applied across concept, pre-vis, filming, and post-production, with examples including Glory, Brasil 70, and The American Experiment.
Users can automate and schedule workflows with ChatGPT, including recurring tasks, and share them with teams. The feature is available now via ChatGPT.
Y Combinator President Garry Tan discusses how AI-native companies achieve 400x productivity leverage, allowing lean teams to operate at scale. He and Eve Bouff capped off the AI Engineer conference's startup and design engineer audiences.
The multi-year collaboration will evaluate how providers' AI use at the point of care correlates with care delivery patterns. The study aims to measure AI's effect on clinical outcomes and efficiency.
Intuit rebuilt its agent architecture twice in four months, VP Nhung Ho said at VB Transform 2026. Ho called the scrapping 'the fast path' to success.
Altman criticized Musk for pitching short-term space data centers to investors, a view most experts already share. Musk responded that his company plans to start flying them next year.
The new feature lets users search with conversational prompts instead of keywords. Rolling out to US users, it aims to improve content discovery on the platform.
Anthropic reset 5h and weekly rate limits on Claude for all users. Users on the $200 Max plan report hitting limits frequently, burning through extra credits, and relying on banked resets. The reset pattern has sparked community tools to track usage windows.
Companies are moving away from reliance on a single AI provider, diversifying models to reduce risk. The trend reflects a strategic shift as enterprises seek flexibility and avoid vendor lock-in.
Suno argues that disclosing its licensing deal with Warner Music would threaten future settlements, responding to Universal Music and Sony Music's push for the contract. The dispute escalates in ongoing copyright litigation.
Databricks introduces a new Context Engineer Certification and agent training programs to address the skills gap in agentic AI. The certification focuses on building and managing AI agents on Databricks' platform.
Over 1 million emails have been found using hidden text (text salting) to evade AI security filters, allowing phishing to slip into inboxes. The technique exploits LLM-based filters that fail to detect invisible characters or hidden content.
Chinese open-weight models accounted for 41% of Hugging Face downloads this spring, surpassing US models. On OpenRouter, six of the top seven most popular models are open Chinese models, with Anthropic's Claude Opus 4.7 in seventh. Hugging Face CEO Clem Delangue noted that enterprises increasingly prefer open models for cost and flexibility, suggesting frontier models may become niche.
The collaboration brings AI models and simulation tools to LeRobot, an open-source robotics platform. It builds on NVIDIA's Cosmos, Isaac, and Jetson platforms to lower barriers for physical AI development.
A survey of 107 enterprises finds AI infrastructure spending accelerating ahead of cost-measurement capabilities. Most run AI on hyperscalers/APIs, but next spend targets specialized compute almost none use today.
Fan-data firm Laylo launches a UGC Agent that automatically collects clips from fans' Instagram Stories for music marketing. It follows their earlier AI Ticket Sales Agent.
DeepSeek is considering a new funding round just weeks after closing a $7 billion round, according to a Financial Times report cited by Bloomberg. No details on size or valuation have been disclosed.
New research finds that language models maintain a shared "value slot" and route beliefs about characters versus reality through it, enabling mental space separation. The mechanism generalizes across beliefs, paintings, memories, and hypotheticals, suggesting a unified computational basis for discourse context.
MUGEN is a comprehensive benchmark for evaluating multi-audio understanding in large audio-language models across speech, general audio, and music. Experiments reveal current LALMs still struggle with multi-audio tasks.