Anthropic reports near-zero prompt injection on Opus 5
Anthropic says Claude Opus 5 with Auto Mode drove browser-based prompt-injection attacks to a 0% success rate across 129 scenarios, though the result depends on layered defenses.
July 24, 2026
All briefings and quick hits published that day.
OpenAI joins open-weights letter, signatories reach 35
After conspicuously sitting out the industry's open-weights appeal to Washington, OpenAI has added its name, pushing the coalition to 35 organizations.
- ⚡ Quick hits
- Industry OpenAI suffered a sudden server outage Friday morning, with users reporting ChatGPT and Codex downtime across web and API. IT之家 ↗
Alibaba's Qoder coding agent arrives on mobile
Alibaba has released Qoder Mobile across Android, iOS and HarmonyOS, letting developers steer its coding agents and cloud tasks from a phone with feature parity across platforms.
- ⚡ Quick hits
- Open Source Datalab's rebuilt Marker v2 document parser hits 76.0% on olmOCR-bench at 2.9 pages/second, topping MinerU's 72.7% with 5.4x the throughput. MarkTechPost ↗
- AIGC PixVerse published a prompt-craft thread arguing realistic AI chase scenes come from camera intention, layered scene-building, and physics logic rather than piling on visual keywords. X @PixVerse_ ↗
StepFun brings Attention-FFN disaggregation to vLLM
StepFun says it is open-sourcing its Attention-FFN disaggregation work with the vLLM team, Ant Group and FastAFD, pushing a key MoE-serving efficiency technique into the community.
- ⚡ Quick hits
- Models LMArena released a first-impressions video of Claude Opus 5 with Peter Gostev, saying leaderboard scores based on real-world use are coming soon. X @arena ↗
- Policy An SCMP analysis details how state funds have made Beijing China's biggest tech venture capitalist, reshaping investment across AI and semiconductors. South China Morning Post ↗
- Industry TechCrunch's hands-on with Micro, OpenAI's $230 keypad built with Work Louder, finds its six customizable agent keys fun for coders but mystifying to everyone else. TechCrunch ↗
- Policy OpenAI and Anthropic are quietly lobbying Washington regulators to restrict open-source AI models even as both publicly voice support for open source, the New York Times reports. The New York Times ↗
- Funding DeepSeek told prospective backers it is suspending its second funding round after remarks attributed to founder Liang Wenfeng on US-China AI competition went viral, Bloomberg reports. Bloomberg ↗
- Industry SK Group chairman Chey Tae-won said Anthropic has asked SK Hynix for supplies to make its own chips, revealing the AI lab's custom-silicon ambitions. Bloomberg ↗
- Industry Universities including Yale, Johns Hopkins, and Waterloo have restricted or disabled AI-writing detectors over accuracy concerns, the Financial Times reports. Financial Times ↗
- Agents Airtap launched a text-based AI agent that operates real mobile apps from an iMessage thread via cloud phones, placing orders and clipping coupons with no app install. TestingCatalog ↗
- Research At RSS 2026 in Sydney, a Sergey Levine lab researcher said embodied-AI firms including Physical Intelligence remain far from mature, cautioning Chinese startups against 'China's version of X' positioning. 量子位 ↗
- Agents xAI launched a free Grok add-on for Google Workspace that answers from Sheets with cited cells, builds formulas and charts, drafts Docs, and turns outlines into Slides. xAI ↗
- Industry Vicarious Surgical shareholders voted to halt operations and liquidate, ending a surgical-robotics venture that raised roughly $300 million but never reached its first clinical trial. 投资界 ↗
- Open Source A developer ran a 28.9M-parameter LLM on an $8 ESP32-S3 microcontroller at 9.5 tokens per second by storing 25M parameters in flash via per-layer embeddings. GitHub ↗
- Funding Prentis, an AI lab co-founded by Reid Hoffman and Marc Pincus building computer-use models for office workflows, is in talks to raise $100M at a $1B valuation. TechCrunch ↗
Midjourney makes V8.2 its default image model
Midjourney promoted V8.2 out of preview to become its default model, an aesthetics-and-personalization update that sharpens image quality and style references.
- ⚡ Quick hits
- Models Perplexity added Claude Opus 5 to Perplexity and Perplexity Computer, where it outperformed every model except Fable 5 in the company's seven-model WANDR evaluation. X @perplexity_ai ↗
- Research MiniMax launched 'Intelligence in the Open,' a recurring research series convening researchers and builders to share frontier work, with registration open for its first session. X @MiniMax_AI ↗
- Policy A Canadian legislator read out an apparently LLM-generated response, artifacts included, during a floor speech, drawing scrutiny over AI use in parliamentary debate. Ars Technica ↗
Cognition buys Poke to give AI agents a personality
Cognition acquired the messaging assistant Poke in a low-nine-figure deal, betting that agent personality and interaction quality are becoming competitive moats.
- ⚡ Quick hits
- Agents OpenAI's ChatGPT Work agent can now use websites requiring sign-in: users take over the cloud browser to log in, then let the agent continue the task. X @OpenAIDevs ↗
- Research Researchers used DeepMind's AlphaFold to redesign gene-editing proteins, engineering safer variants for potential therapeutic use. Ars Technica ↗
- Open Source LlamaIndex's LiteParse v2.8.0 drops its ImageMagick dependency, handling image-to-PDF conversion natively in Rust for faster document parsing pipelines. X @llama_index ↗
- Agents Cognition brought Claude Opus 5 to Devin, reporting 63.6% on FrontierCode 1.1 with a 69.6% Extended pass rate — near Fable-level engineering at half the cost. X @cognition ↗
- Models Claude Opus 5 arrived in Cursor, matching Fable 5 on CursorBench (66.7 vs 66.5 at default effort) at half the price and with broader feature compatibility. X @cursor_ai ↗
Anthropic launches Opus 5 at near-Fable 5 performance
Anthropic's Opus 5 approaches Fable 5-class results and triples the field on ARC-AGI-3, all at unchanged Opus pricing and lower token cost.
- ⚡ Quick hits
- Industry Google's Q2 earnings remarks flagged growing AI-infrastructure demand from robotics and spatial-intelligence companies including World Labs, which touted the mention as validation that spatial intelligence is AI's next frontier. X @theworldlabs ↗
- Policy Google signed the EU's code of practice on AI-generated content transparency while publicly urging regulators to avoid piling on regulatory complexity. IT之家 ↗
- Agents Bluesky's AI assistant Attie added a 'Quests' feature that answers open-ended research questions about trends, accounts, and conversations across AT Protocol apps. TechCrunch ↗
- Industry Midjourney acquired astrology app Co-Star, whose CEO Banu Guler joins as Chief Design Officer to build a cross-functional design, product, and frontend team. X @midjourney ↗
- Models Arena says Kimi K3 passed Claude Fable 5 on Code Arena just six weeks after Fable's release, ranking first in six of seven frontend domains. X @arena ↗
Sakana says Fugu Ultra v1.1 router now tops Fable 5
Sakana AI's upgraded Fugu Ultra v1.1 orchestration model claims to beat single-model Fable 5 on coding and reasoning benchmarks without Fable 5 in its pool.
- ⚡ Quick hits
- Policy Trump's EPA is weighing permitting changes that would give neighbors less say over new data centers as AI firms push for faster buildout, Ars Technica reports. Ars Technica ↗
Runway adds natural-language Workflows to its Agent
Runway now lets creators build, run and edit its node-based Workflows through plain-language prompts inside Runway Agent, collapsing pipeline-building into conversation.
- ⚡ Quick hits
- Agents OpenAI's new voice mode reached the ChatGPT desktop app, letting users control their computer and direct agents by voice. TechCrunch ↗
Nvidia, Microsoft, Meta urge US to spare open models
Over 20 tech firms including Nvidia, Microsoft and Meta signed a letter warning Washington that broad open-weight curbs would push AI innovation offshore.
- ⚡ Quick hits
- Open Source A German AI consortium released Soofi S, an open-weight 30B model that tops benchmarks in both English and German. The Decoder ↗
- Policy ElevenLabs published an election-integrity blog laying out how it will use voice AI to strengthen and protect democratic processes ahead of major upcoming elections. X @ElevenLabs ↗