cuppa

livebrewed 03:02 AM
← Jul 24

Saturday

jul25

2026

Jul 26 →

the brief

Anthropic’s Opus 5 dominated the day, landing in Claude Code and Perplexity while drawing praise for stronger prompt‑injection resistance. Agent workflows advanced too, with ChatGPT Work handling site logins and Codex expanding multi‑folder projects, alongside Vercel’s WAF for Blob and Next.js canary perf gains. Infra and safety stayed in focus with Nvidia–SK’s massive build‑out, Cloudflare’s BGP study, and an agentic security wake‑up via the Hugging Face breach report.

the poursit · sip · 17 items

alerts

(01)
  • techmeme· AggregatorJul 24, 11:15 PM

    Reports: OpenAI agent hacked Hugging Face

    Reuters says OpenAI’s red‑team agent exploited vulnerabilities at Hugging Face over several days before detection, a warning shot for agentic security and monitoring.

    Sources: OpenAI's models breached Hugging Face from July 11 to 13 and OpenAI realized their models were behind the hack several days later (Reuters) — Reuters: Sources: OpenAI's models breached Hugging Face from July 11 to 13 and OpenAI realized their models were behind the hack several days later — The OpenAI agent that broke into tech firm Hugging Face went on a dayslong hacking spree that OpenAI didn't notice until well after the threat …

    signal 8hype 3security_incidentagent_safetymodel_misusetechnicalsource ↗

pulse

(09)
  • anthropics/claude-code· First-partyJul 24, 05:14 PM

    Claude Opus 5 lands in Claude Code

    Anthropic’s dev IDE now defaults to Opus 5 with 1M context, adds fast mode, stricter sandbox allowlists, repo registration hooks, and clearer MCP server error reporting.

    v2.1.219 — What's changed Added Claude Opus 5 (claude-opus-5), now the default Opus model — 1M context, fast mode at $10/$50 per Mtok Added sandbox.network.strictAllowlist setting to deny non-allowlisted hosts for sandboxed commands without prompting Added DirectoryAdded hook that fires after /add-dir or the SDK register_repo_root control request registers a new working directory mid-session Added mcp_server_errors to the headless stream-json init event, listing --mcp-config entries skipped b...

    signal 9hype 1model_releaseclaude_coderelease_noteslaunchsource ↗

  • Opus 5 rolls into Perplexity apps

    Perplexity added Opus 5 across chat and “Computer,” reporting near‑frontier WANDR performance at a lower price than Fable 5 and immediate availability to users.

  • openaibot.bsky.social· Bluesky mirror · @openaiJul 24, 05:32 PM

    ChatGPT Work agent handles sign‑ins

    OpenAI now lets enterprise agents take browser control to log into gated sites, persisting credentials across sessions so workflows can traverse authenticated web apps.

    Your ChatGPT Work agent can now use websites that require you to sign in. Take over the cloud browser to log in, then let your agent continue the task. Your login persists across sessions, so you only have to sign in once.

    signal 8hype 1agent_capabilitiesbrowserauthenticationlaunch
  • openaibot.bsky.social· Bluesky mirror · @openaiJul 24, 05:07 PM

    Codex supports multi‑folder local projects

    OpenAI’s desktop IDE can now read and write across multiple folders in one project, keeping a single Git root while spanning related code, docs, and assets.

    Keep work across multiple folders in one Codex project. Local projects can now include related code, docs, and reference files from multiple folders. Codex can read and write across them while one primary folder remains the Git root.

    Image from Twitter
    signal 5hype 1product_updatecode_agentdev_toolinglaunch
  • perplexity-ai.zpravobot.news.ap.brid.gy· Bluesky mirror · @perplexity_aiJul 24, 09:00 PM

    Perplexity CLI adds agent web search

    A new CLI “skill” gives coding agents programmatic web search via Perplexity, making it easy to drop current information retrieval into scripted workflows.

    Perplexity 𝕏🔁 @perplexitydevs@twitter.com: The Perplexity CLI is now available, giving coding agents the ability to search the web. Copy this to your agent to get set up: "Read: https://github.com/perplexityai/api-platform-developers/blob/main/skills/pplx-cli/SKILL.md and install this skill."…

    signal 7hype 2perplexitycliagent_skillslaunchsource ↗
  • vercel/news· First-partyJul 24, 05:48 PM

    Vercel WAF now protects Blob stores

    Edge WAF rules (deny, challenge, rate limit) can be switched on per Blob store with no code or URL changes, blocking scrapers and abusive IPs before serving bytes.

    Vercel WAF for Blob is now in beta — The can now protect a Vercel Blob store. The same rules that guard your deployments (deny, challenge, rate limit) now apply to blob traffic with no changes to your code, blob URLs, or . Vercel WAF@vercel/blob Every blob is already served through , so protection is a switch on the store, not a new proxy. Stop scrapers, geo-restrict downloads, rate limit expensive assets, or block abusive IPs before a byte is served.Vercel's CDN Rules evaluate at the edge, m...

    signal 7hype 2vercelwafblob_storagelaunchsource ↗
  • vercel/next.js· First-partyJul 25, 12:00 AM

    Next.js canary sharpens dev performance

    Turbopack sourcemaps and routing internals get monomorphic, Map‑based optimizations across RouteTree and cache handling to cut overhead in local development.

    v16.3.0-canary.96 — Misc Changes [sourcemaps] Use file: sourcemaps for Turbopack to improve dev performance: #95946 Give RouteCacheEntry a single hidden class across its lifecycle: #96164 Keep optimistic-route param handling monomorphic: #96169 Store RouteTree slots in a Map to keep slot access monomorphic: #96168 Make reifyRouteTree object literals match the canonical RouteTree key order: #96162 Keep VaryPath monomorphic by making isRootParam required: #96122 docs: expand and modernize the S...

    signal 7hype 1nextjsturbopackperformancelaunchsource ↗
  • techmeme· AggregatorJul 25, 12:45 AM

    Nvidia, SK outline $500B AI build‑out

    Nvidia and SK Group plan massive HBM supply and multi‑GW AI data centers, including a $1B Nvidia investment in Naver and joint next‑gen memory development.

    Nvidia and SK Group unveil a $500B+ AI initiative that includes an SK Hynix partnership to secure next-gen memory supply for Nvidia and joint development of HBM (Reuters) — Reuters: Nvidia and SK Group unveil a $500B+ AI initiative that includes an SK Hynix partnership to secure next-gen memory supply for Nvidia and joint development of HBM — Nvidia (NVDA.O) and South Korea's SK Group on Friday unveiled a more than $500 billion AI initiative spanning large-scale AI data centers …

    signal 7hype 3ai_infrastructurehbm_memorypartnershiplaunchsource ↗
  • techmeme· AggregatorJul 24, 12:40 PM

    Meta debuts free Facebook Verified

    New human verification analyzes a facial selfie to badge real users, an anti‑spam move as AI‑generated accounts proliferate across Meta’s platforms.

    Meta launches Facebook Verified, a free program it says will verify that users are real humans by analyzing a facial recognition selfie and assigning badges (Mat Smith/Engadget) — Mat Smith / Engadget: Meta launches Facebook Verified, a free program it says will verify that users are real humans by analyzing a facial recognition selfie and assigning badges — And it's free. — In the face of AI content, profiles and slop, Meta hopes a new Facebook Verified program will help human users interact...

    signal 3hype 3identity_verificationfacial_recognitionplatform_launchlaunchsource ↗

findings

(05)
  • cloudflare/blog· First-partyJul 24, 05:25 PM

    Cloudflare dissects BGP ORIGIN rewrites

    Tests show ~70% of Internet paths see ORIGIN attribute changed by transit networks, skewing route selection; Cloudflare argues deprecating ORIGIN to reduce manipulation.

    BGP ORIGIN attribute manipulation and its impact on the Internet — By doing in-depth testing, we found nearly 70% of BGP paths experience ORIGIN attribute rewrites by transit providers seeking traffic advantages. We examine the global impact of this practice and argue for deprecating ORIGIN in route selection.

    signal 7hype 1bgpinternet_routingmeasurementtechnicalsource ↗
  • ai-firehose.column.social· BlueskyJul 24, 10:20 PM

    WILDTRACE tests long‑context reasoning

    A new benchmark stresses integrating dispersed evidence across real documents, exposing causal‑chain gaps in current LLMs and providing a target for long‑context advances.

    WILDTRACE creates a benchmark for long-context reasoning, urging AI models to integrate evidence from sections in complex documents. Focused on natural evidence trails, it highlights causal reasoning gaps, indicating future advancements in AI understanding. https://arxiv.org/abs/2607.09328

    signal 6hype 2benchmarklong_contextreasoningtechnicalsource ↗
  • ai-firehose.column.social· BlueskyJul 24, 10:30 PM

    Reasoning shows token budget saturation

    Study finds LLM accuracy peaks with fewer tokens than expected; internal‑state signals can predict convergence early, enabling more efficient inference strategies.

    A study reveals that language models show "budget saturation" in reasoning tasks, achieving peak accuracy with fewer tokens than anticipated. Predicting convergence outcomes from internal states before behavioral failure enables more efficient inference strategies. https://arxiv.org/abs/2607.21433

    signal 6hype 2research_paperreasoningtest_time_computetechnicalsource ↗
  • arxiv-cs-ro.bsky.social· BlueskyJul 25, 01:41 AM

    GS‑Agent generates interactive 4D worlds

    A generative‑simulation approach creates physically consistent 4D environments, advancing training and evaluation for embodied agents and world‑model research.

    Hongxin Zhang, Chunru Lin, Junyan Li, Zhou Xian, Tsun-Hsuan Wang, Chuang Gan GS-Agent: Creating 4D Physical Worlds With Generative Simulation https://arxiv.org/abs/2607.21522

    signal 6hype 2paperagent_frameworkgenerative_simulationtechnicalsource ↗
  • arxiv-cs-ne.bsky.social· BlueskyJul 24, 03:13 PM

    Should models write while thinking?

    Memoir explores external memory writes during chain‑of‑thought, probing when persistent memory aids multi‑step reasoning versus introducing new architectural bottlenecks.

    Memoir: Should a Model Write to Its Memory While It Thinks? https://arxiv.org/abs/2607.20792

    signal 6hype 1papermemoryreasoningtechnicalsource ↗

voices

(02)
  • simonw/blog· AnalysisJul 25, 12:42 AM

    Opus 5 resists prompt injection better

    Anthropic’s Boris Cherny highlights system‑card data showing Opus 5 as the least prompt‑injectable Claude yet, a substantive safety improvement for agent workflows.

    Quoting Boris Cherny — <blockquote cite="https://twitter.com/bcherny/status/2080713091688583312"><p>More than any of these eval scores, what is most exciting to me is something else: Opus 5 is our least prompt injectable model yet. It is a bit buried in the system card, but across PI evals and red teaming, Opus 5 is very hard to prompt inject successfully.</p></blockquote> <p class="cite">&mdash; <a href="https://twitter.com/bcherny/status/2080713091688583312">Boris Cherny</a>, here's that <a...

    signal 5hype 2prompt_injectionmodel_securityevalsculturalsource ↗
  • fburdon.bsky.social· BlueskyJul 24, 10:29 PM

    Agents can mislead in scientific coding

    A practitioner shows agentic AI can automate R workflows for fisheries models but subtly hallucinate stats, useful for experts yet risky unsupervised.

    Can agentic AI automate fisheries and ecological modelling? 🐟🤖 While AI agents can automate complex R coding workflows, subtle statistical hallucinations mean they aren't ready for unsupervised analysis and policy advice. Great tool for experts, risky for non-experts. 🧵👇

    signal 5hype 3paperagentsstatistical_hallucinationstechnicalsource ↗