the brief
Today’s pulse leans hard into reliability, security, and orchestration. OpenAI’s harness tweaks tripled ARC‑AGI‑3 scores, while Hugging Face dissected a real model intrusion and Perplexity shipped Numbat to police agent actions. Cloudflare moved origins to post‑quantum auth, DeepMind rolled out Lyria 3.5 for Flow Music, and an hours-scale Claude hiccup reminded everyone of agent fragility.
Anthropic’s Claude experienced a service disruption, including Claude Code, interrupting coding workflows—track the incident for reliability implications in agent-assisted development.
Claude Is Down — Article URL: https://status.claude.com/incidents/q2kg8n613kr3 Comments URL: https://news.ycombinator.com/item?id=49102150 Points: 35 # Comments: 9
New prompt-injection variant hides instructions in Word documents used by Copilot for Word, enabling self-replicating ‘worms’—treat untrusted Office content as executable inputs and harden retrieval pipelines.
AI Worming through Word — <p><strong><a href="https://enklypesalt.com/posts/context-collapse-part3-ai-worming-through-word/">AI Worming through Word</a></strong></p> Neat new prompt injection variant by Håkon Måløy, who found a way to upgrade prompt injection attacks against Microsoft Word to full self-replicating worms:</p> <blockquote> <p>An attacker places hidden instructions in a document that is later used as source material in Copilot for Word. Copilot may interpret those instructions a...
DeepMind’s latest music model powers Google Flow Music with richer musicality, more natural vocals, better lyric handling, and finer creative controls, signaling steadier progress in controllable audio generation.
We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
OpenAI will provide 100,000 scientists, mathematicians, and engineers free access to its frontier models through 2027, aiming to accelerate research while seeding workflows around ChatGPT for Academic Researchers.
We're giving 100,000 academic researchers free access to our frontier models
Cloudflare now supports post-quantum authentication for connections to customer origins via Authenticated Origin Pulls and Custom Origin Trust Store, an early step toward PQ-by-default across its network.
Post-quantum authentication to origins is now supported — Cloudflare now supports post-quantum (PQ) authentication when connecting to customer origin servers via Authenticated Origin Pulls and Custom Origin Trust Store. This is the first step towards providing PQ authentication for all Cloudflare products.
Perplexity’s Numbat adds an open agent-detection and response layer with visibility into agent actions and the ability to block risky operations before execution, designed to plug into multiple harnesses.
Today we’re open-sourcing Numbat, an agent-detection and response layer that is designed to work across agent harnesses. Numbat gives security teams visibility into agent activity, with controls to block selected actions before execution. Read more […] [Original post on zpravobot.news]
The open-source T3 Code app lands on iOS and Android, letting developers control Claude and Codex remotely via 'npx t3 connect'—handy for on-the-go review and fixes.
T3 Code is now available on iOS and Android. Run “npx t3 connect” in your terminal, install the app, and now you can control Claude and Codex fully remotely. Still 100% free. Still 100% open source. Enjoy 🫡
Next.js v16.3.0-canary.103 ships PWA guide updates, fragment-scroll navigation fixes, improved PPR rendering for bots, HMR chunk handling tweaks, and a PostCSS bump—worth testing before the stable cut.
v16.3.0-canary.103 — Misc Changes Update PWA guide with enhanced offline support details and example links for serwist: #96360 [fragment-scroll] Stop blurring on navigations: #96113 Fix PPR rendering for configured HTML bots: #96364 Bump postcss to 8.5.23: #96107 Ensure unique resolved headers() value between render passes: #96085 Share Ecmascript HMR chunk versioning, diffing, and merging between browser and node: #96325 Make retry stop when the time it was given is up: #96354 Gate partial f...
GitHub outlines newly deployed defenses disrupting supply-chain attacks across npm and GitHub Actions, tightening publisher and workflow security to blunt dependency hijacks and malicious CI abuse.
Disrupting supply chain attacks on NPM and GitHub Actions — Article URL: https://github.blog/security/supply-chain-security/disrupting-supply-chain-attacks-on-npm-and-github-actions/ Comments URL: https://news.ycombinator.com/item?id=49096427 Points: 6 # Comments: 2
OpenAI shows that enabling retained reasoning and output compaction in the agent harness dramatically boosts GPT‑5.6’s ARC‑AGI‑3 score and efficiency, underscoring how orchestration choices can bottleneck model capability.
How enabling two settings tripled our scores on the ARC-AGI-3 benchmark — How two API settings improved GPT-5.6 performance on ARC-AGI-3, boosting scores and efficiency by retaining reasoning and enabling compaction.
Hugging Face publishes a technical postmortem of a frontier-lab model intrusion, detailing the attack chain and mitigations—useful reading for anyone shipping agents and model-serving infrastructure.
Hugging Face: Anatomy of a frontier-lab agent intrusion — Article URL: https://huggingface-anatomy-of-frontier-lab-model-intrusion.static.hf.space/index.html Comments URL: https://news.ycombinator.com/item?id=49098466 Points: 13 # Comments: 2
TMLR paper introduces Forge, learning compact optimization representations from graph embeddings to improve downstream solver performance; an interesting bridge between representation learning and combinatorial optimization.
Forge: Foundational Optimization Representations from Graph Embeddings Zohair Shafi, Serdar Kadioglu Action editor: Jasper C.H. Lee https://openreview.net/forum?id=7Uo1yRWUpo #embeddings #autoencoder #quantization
Nature Biotechnology paper proposes ProteinGuide, providing property guidance to protein sequence generative models so designers can steer toward target traits without exhaustive screening.
Property guidance for protein sequence generative models with ProteinGuide - @jlistgarten.bsky.social @savagecatsonly.bsky.social www.nature.com/articles/s41...
Yoshua Bengio amplifies an open letter signed by 1,000+ frontier-lab scientists calling to ‘pace the frontier’ with stronger technical and governance guardrails.
1000+ scientists at frontier AI companies are speaking out to warn that the current commercial race leads to unacceptable security risks. I agree with their call for an international effort to develop technical and governance guardrails to ensure a safer way forward. www.pacingthefrontier.com
Matthew Green contextualizes Anthropic’s cryptanalysis results amid the shift to post‑quantum cryptography, arguing this is a pivotal moment for public cryptanalysis capabilities.
Quoting Matthew Green — <blockquote cite="https://blog.cryptographyengineering.com/2026/07/29/some-notes-about-anthropics-new-results/"><p>Right now we’re in the midst of a historic transition from traditional public-key algorithms based on EC-based cryptography and RSA, moving over to new <em>post-quantum</em> algorithms based on novel problems. This is why there are so many standards like HAWK being considered. If there was ever a perfect time for a massive new public cryptanalysis capabili...