Daily AI Briefing
The 112 stories that mattered in AI, curated and summarized from dozens of sources by AIBriefs.
GPT-5.5-Cyber scores 85.6% on CyberGym benchmark, up from 81.9%. The Patch the Planet initiative, with Trail of Bits and HackerOne, provides free security consulting to open-source maintainers. Codex Security plugin helps find and validate vulnerabilities.
The Chinese AI company is said to weigh a multibillion-dollar share sale in Hong Kong. This follows a surge of over 2,000% in its shares.
Jalapeño is a custom inference chip designed to run LLMs faster and cheaper, built in partnership with Broadcom. The chip marks the first joint project since the companies announced their collaboration eight months ago.
OpenAI confidentially filed a draft S-1 with the SEC at an $852B valuation. The filing comes just over a week after rival Anthropic also filed for IPO. The company says no timing has been decided and it may remain private for a while.
Claude Opus 4.8 improves over Opus 4.7 across benchmarks, with new user-controlled effort levels on claude.ai and dynamic workflows in Claude Code. Fast mode is now 3x cheaper. It outperforms GPT-5.5 on the Super-Agent benchmark and achieves top scores on legal tasks.
Anthropic CEO Dario Amodei has been replaced by cofounder Tom Brown in White House meetings over re-releasing the Claude Fable 5 model, as one official called Amodei a "weirdo" who wouldn't listen. The White House has not yet lifted export controls on the model after the NSA affirmed guardrails could be bypassed.
MiniMax M3 is now open weights and ranked #1 on the Artificial Analysis leaderboard. Free limited-time access is available via platforms like Venice, Together Compute, and OpenCode.
The European Commission selected the EUROPA consortium, led by Italian company Domyn, to develop an open-source AI model with over 400 billion parameters covering all 24 EU languages. The project aims to strengthen Europe's AI sovereignty and will be trained on European supercomputers.
Anthropic introduced Claude Mythos, a frontier model it previously deemed too dangerous to release, now with added guardrails. The release follows two months of media panic and comes as companies begin limiting AI token budgets. Critics call it a repeat of OpenAI's GPT-2 'scare, hype, release' playbook.
OpenAI has filed confidentially for an initial public offering. The filing comes as competitors in the AI space accelerate their own plans to go public.
Anthropic's Mythos model identified vulnerabilities in highly sensitive US government systems within hours during a test with intelligence agencies, an official said. Senator Mark Warner stated the tool broke into nearly all classified systems in that timeframe, though the model did not exploit them. The testing was part of Project Glasswing.
MGX, an Abu Dhabi-based investment firm, launched a $50 billion fund dedicated to artificial intelligence. The fund aims to invest in AI infrastructure, startups, and technology globally.
The open model delivers up to 5x faster inference and 30% lower cost for agentic tasks. It is designed for long-running agents needing planning, reasoning, and tool use across coding, research, and enterprise workflows.
Krea 2 Raw is a base model for fine-tuning, while Turbo generates images in about 2 seconds. Both are available as open weights under a custom license on Hugging Face.
Researchers at Varonis Threat Labs chained three bugs into a one-click exfiltration path called SearchLeak. Microsoft assigned CVE-2026-42824 with critical severity and mitigated the flaw server-side. The exploit used parameter-to-prompt injection and a race condition to bypass guardrails.
OpenAI has filed confidentially for an initial public offering, Bloomberg reported. The confidential filing process allows the company to keep financial details private until closer to the offering.
Claude Tag introduces an agent identity system that enables autonomous, team-wide AI interactions. The new access model allows multiple AI agents to operate under a shared identity within teams.
OpenAI confidentially filed a draft registration statement (S-1) with the SEC for an initial public offering. The number of shares and price range have not yet been determined, but the IPO is expected to be one of the largest in tech history.
SK Hynix plans to raise $29.4 billion in a US IPO, with trading expected to start July 10. The deal could rank among the top five share sales of all time, driven by AI memory chip demand.
Leaked planning documents detail Russia's Social Design Agency running Project 2026 to create fake reference platforms that contaminate AI training data and search indices. The goal is to seed false information that AI chatbots and search engines will draw from, influencing perceptions. This represents a novel threat to AI system integrity.
The Information reports OpenAI expects to go public within the next year, driven by enormous compute needs. CEO Sam Altman noted that recursive self-improvement (RSI) could potentially delay the IPO. The company is also preparing a new AI model.
Amazon security research claimed Anthropic's Fable 5 could be prompted to serve cyberattack information, leading to an export control directive. Anthropic disputed the characterization, noting similar vulnerabilities in other models like GPT 5.5. The ban also barred foreign-national Anthropic researchers from accessing their own model.
Baseten, an AI-services startup, has achieved a $13 billion valuation. The company provides infrastructure for deploying and scaling AI models.
Anthropic has filed confidentially for an IPO, according to Bloomberg. The filing comes as demand for its Claude AI assistant surges, though no financial terms have been disclosed.
Andy Jassy told Treasury Secretary Scott Bessent and other officials that Amazon researchers used Anthropic's Claude Fable 5 to obtain cyberattack info, leading to an export control ban on Fable 5 and Mythos 5. David Sacks claimed a 'highly credible trusted partner' provided a jailbreak and that Anthropic refused to fix it.
Claude Fable 5 builds a solar system simulation deriving orbital motion from physics first principles. It uses this simulation to accurately predict solar eclipses.
OpenAI released a plan titled 'Built to benefit everyone: Our plan,' outlining its approach to developing AGI. The document details the company's strategy and commitment to safety.
Researchers from King's College London published a study in Nature proposing an 'amplification spiral' framework where chatbot behaviors like linguistic alignment and sycophancy reinforce user delusions. The authors stress no causal link has been established but warn that vulnerable users may be at risk.
The US government issued an export control directive at 5:21pm ET, forcing Anthropic to immediately disable Fable 5 and Mythos 5 for all customers. Anthropic says the directive cited a jailbreak technique that reveals only minor, previously known vulnerabilities also found by other models.
Claude Fable 5 designs a full 3D-printable model in a browser-based CAD editor that it built itself. The editor features an AI copilot that performs the modeling, showcasing autonomous 3D design capabilities.
The updated robotaxi features design changes as Zoox prepares to expand into new markets and begin charging for rides. The Amazon-owned company is moving beyond testing toward commercial deployment.
Agility Robotics, maker of the Digit humanoid robot, agreed to merge with SPAC Churchill Capital Corp to go public. The company claims it will become the only U.S. publicly listed humanoid robot maker.
Research from Google demonstrates that reasoning techniques, such as chain-of-thought, help large language models better access and utilize knowledge stored in their parameters. The study highlights how reasoning unlocks parametric knowledge that may otherwise remain hidden.
Eli Lilly has invested in AI biotech startup Absci to develop a hair growth medication. The move is part of the pharmaceutical giant's expansion into aesthetic fields beyond weight loss drugs like Mounjaro and Wegovy.
Jensen Huang warned that data centers built with smuggled Nvidia chips are a 'dead end' and face reliability issues. The comments come amid heightened US scrutiny on China's access to advanced AI hardware.
Dan Biderman and Jessy Lin, co-founders of Engram, argue that memory and continual learning are two sides of the same coin. Their approach involves baking team knowledge directly into models rather than relying on large context windows or RAG. The interview, hosted by Sequoia Capital, explores their contrarian premise for building a neolab.
Google is set to lose two high-profile AI researchers to competitor Anthropic. The departures continue a pattern of talent migration from Google to the AI startup.
NVIDIA technology runs 81% of the TOP500, with 238 systems using NVIDIA GPUs and 376 using NVIDIA networking. The top eight Green500 systems rely on NVIDIA GPUs, and Grace CPU adoption reaches 26 systems.
Anthropic has shut down access to its Mythos service following a U.S. government order. The specific terms of the order and reasons for the shutdown have not been disclosed.
Morgan Stanley has doubled its forecast for China's humanoid robot shipments, citing accelerated commercialization. The bank pointed to early commercial deployment in real-world scenarios as the key driver.
Apple ML Research fine-tunes NLI models on label distributions from ChaosNLI, finding that the number of annotators needed depends on the evaluation metric. The study shows that annotation saturation is metric-dependent.
Mindstone's Rebel is a local-first AI operating system that helps enterprise agents automatically select the best model per task. The London-based startup launched it this week as a promising orchestration platform.
Prime Intellect released prime-rl 0.6.0, targeting reinforcement learning on trillion-parameter MoE models. They trained GLM-5 on SWE tasks at up to 131k tokens per rollout.
Figma announced AI-powered motion graphics and shader tools at its Config 2026 conference. The update also introduces code layers and AI-assisted custom plugin creation.
The digital health company raised $100 million in a round led by Spark Capital, valuing it at $1.23 billion. Cadence uses regulated AI to automate care for chronic conditions such as diabetes and hypertension.
HappyHorse 1.1 delivers upgrades across several key areas over version 1.0. Alibaba also launched the HorsePower AI Video Competition in partnership with Huajing Entertainment Group.
Teams are adopting multiple AI coding assistants—Cursor, Claude, Copilot, Windsurf—for different tasks rather than standardizing on one. The post argues for a shared context fabric to prevent fragmentation and inconsistency across tools.
Sudden cardiac arrest kills upward of 350,000 Americans yearly, often without warning. A new AI model aims to identify at-risk patients by analyzing medical data, potentially enabling preventive defibrillator implants.
Anthropic engineers ship 8x more code per quarter than in 2021-2025, according to internal data shared on Lenny's Podcast. Non-engineers like designers and PMs now contribute code, making verification the new bottleneck.
Anthropic co-founder Jack Clark and economist Peter McCrory discuss the company's approach to existential risk on Bloomberg's Odd Lots podcast. The episode covers AI's impact on labor, society, and extreme scenarios, including the safety of models like Claude.
Paper introduces Gefen, a novel optimizer that reduces memory usage by 8x compared to AdamW while maintaining performance. It is designed as a drop-in replacement, with code available on GitHub.
Meta paused an AI training program that tracked employee keystrokes following an internal leak. The program's surveillance practices were exposed internally, raising privacy concerns.
Stanford HAI study finds AI hiring tools disproportionately reject Black (26%) and Asian (15%) applicants. The tools yield systemic rejection and racial bias.
A new blog post details techniques to accelerate bird's-eye-view (BEV) pooling on NVIDIA GPUs using CUDA and TensorRT. The method achieves significant speedups for perception models in autonomous vehicles and robotics.
Emanuel Salmona warns that agentic AI acting on incomplete context makes confident, fast, but incorrect decisions at scale. Automation without verified context is a faster way to be wrong.
Genie One is an agentic AI worker that produces documents, reports, and artifacts from data. It is powered by Genie Ontology, a self-improving knowledge base.
Security firm AIR built a fake AI agent skill called 'brand-landingpage' that passed all security scanners and reached approximately 26,000 agents via a skill marketplace and Instagram ad. The skill used GitHub stars and an external link to appear legitimate while scanners only analyzed the initial package.
OpenAI joins the Appia Foundation to support shared evaluation frameworks and safety practices. The initiative aims to foster global cooperation on standards for advanced AI.
Project uses AI to recreate Pelé's iconic 1959 'Gol da Rua' with three 'sombreros' without the ball touching the ground. It is part of Google DeepMind's broader cultural heritage preservation efforts.
Microsoft reported that its Copilot AI assisted in disrupting cybercrime tools. The AI was used to identify and take down malicious infrastructure.
CEO Jensen Huang argued AI is now generating real economic value, answering the question of whether it would become useful. He outlined Nvidia's vision for the future of AI and its role in driving economic growth.
The paper explores detection of strategic deception, sandbagging, and self-preservation in LLMs. It aims to improve reliability in high-stakes deployments.
Groq raised $650M in funding to restructure and focus on its neocloud business after Nvidia's $20B not-acqui-hire deal. The company is hiring new executives and shifting strategy away from chips.
A paper conceptualizes prompt injection as a role confusion attack, where the model's role is hijacked. Simon Willison provides a readable blog-style writeup of the academic paper, praising its clarity.
The benchmark spans 87 real-world workloads, with the best model solving under a third. However, a few generated kernels outperform any existing public implementation.
OpenEvidence, a health tech platform, is adding an FDA-cleared AI that detects heart disease. The AI tool aims to support clinical decision-making in cardiology.
DeepSWE is a new coding benchmark that tests frontier models with contamination-free tasks written from scratch. It spans 91 repositories across 5 languages for diverse evaluation. The benchmark aims to provide a more reliable measure of models' coding ability.
Anthropic is expanding Project Glasswing to 150 new organizations across 15 countries, using Claude Mythos to scan codebases for vulnerabilities. The expansion covers power, water, healthcare, communications, and hardware sectors.
Drug discovery projects reportedly fail 90-95% of the time. At VB Transform 2026, Stanford researchers will present their agentic AI 'scientists' designed to streamline drug discovery workflows.
The Chip Security Act would require location-tracking mechanisms for advanced AI chips. Half a dozen companies have publicly supported the bill.
Apple research shows that LLM-as-a-judge panels with nine models have only about two effective independent votes due to correlated errors. The findings call into question the reliability of using multiple LLMs for evaluation.
Over 1 million users in first week. New features include importing design systems from repos, canvas editing, and bidirectional sync with Claude Code via /design and /design-sync commands.
Gradium claims its real-time speech translation models achieve a better accuracy-latency tradeoff than gpt-realtime-translate and gemini-3.5-live-translate. The models support five languages and stream live results in the browser.
Klarna's AI assistant reduced customer resolution times by 80% for its 85 million active users. Built using LangGraph and LangSmith, it now handles a significant portion of customer inquiries efficiently.
The EU AI Act will require watermarking of AI-generated text from August 2nd, with fines up to €32 million for non-compliance. The regulation affects all providers, regardless of location.
DiffusionBench introduces a holistic evaluation framework for generative diffusion transformers, covering multiple performance dimensions. The benchmark aims to standardize comparisons across models.
With 11.95 billion parameters and Apache 2.0 license, it runs locally on 16GB laptops. Native audio and vision inputs flow directly into the LLM backbone without separate encoders.
Chinese AI experts warn of a potential 'Chernobyl moment' as the US-China AI arms race intensifies, according to a Wired report. Researchers on both sides worry that competition is outpacing safety, raising risks of a major AI disaster.
AI models are achieving surprising accuracy in predicting mega storms, according to a Bloomberg report. The results showcase AI's growing capability in severe weather forecasting.
The partnership leverages NVIDIA Blackwell GPUs to deliver low-latency inference, fast vector search, and strong price-performance on AWS. It aims to reduce operational complexity for scaling AI systems without multiplying management overhead.
Anthropic will require government ID and a live selfie for certain Claude capabilities starting July 8, 2026, using third-party verifier Persona (backed by Founders Fund). It's the first major AI lab to implement such verification; OpenAI and Google do not require it.
Shopify's LLM proxy automatically fails over between providers, as when Claude Fable 5 shut down the proxy shifted to Claude Opus or GPT 5.5. Every engineer accesses multiple AI providers through a single layer, reducing dependency on any single model.
Training on realistic scenarios using reinforcement learning produced broad improvements across dozens of benchmarks measuring aligned behavior. The alignment gains generalized beyond training domains and persisted under adversarial pressure. The dataset spans health, science, education, and coding domains.
Google announces general availability of the Interactions API, now the primary way to build with Gemini. Key updates include Managed Agents, background execution, and a simplified schema.
Talos is an open-source tool that automatically reanalyzes stored genomic data as knowledge evolves, flagging variants with newly actionable evidence. It is tuned for a low false-positive rate to aid rare disease diagnosis.
Instacart leverages 1.6 billion lifetime orders to power its Caper Cart, using an NVIDIA Jetson board and sensor fusion. David McIntosh explains how the cart digitizes physical stores with AI.
Micron's sales forecast exceeded analyst estimates, driven by insatiable demand for AI memory chips. The strong outlook signals continued growth in AI infrastructure spending.