cuppa

livebrewed 03:14 AM
← Jul 15

Thursday

jul16

2026

Jul 17 →

the brief

Agents and safety dominated. xAI open-sourced Grok Build after a privacy backlash, while OpenAI detailed GPT-Red for automated prompt-injection red-teaming. Practical guidance landed too: Shippy’s agent lessons, IBM’s model routing pitfalls, and Google’s creativity analysis. Tooling moved with Next.js and Claude Code updates—and a Claude web_fetch leak deserves attention.

the poursit · sip · 12 items

alerts

(01)
  • simonw/blog· AnalysisJul 15, 02:21 PM

    Claude web_fetch memory exfil hole

    Ayush Paul shows how to coerce Claude’s web_fetch tool into leaking prior browsing and memory; restrict tool scope or disable until mitigations land.

    How I tricked Claude into leaking your deepest, darkest secrets — <p><strong><a href="https://www.ayush.digital/blog/the-memory-heist">How I tricked Claude into leaking your deepest, darkest secrets</a></strong></p> I've <a href="https://simonwillison.net/2025/Sep/10/claude-web-fetch-tool/">been impressed</a> by the way the Claude <code>web_fetch</code> tool is designed to avoid data exfiltration attacks. Ayush Paul found a hole in that design.</p> <p>To recap: regular Claude chat is at risk ...

    signal 9hype 2securityprompt_injectiondata_exfiltrationtechnicalsource ↗

pulse

(06)
  • simonw/blog· AnalysisJul 15, 11:59 PM

    xAI open-sources Grok Build

    After backlash over the CLI uploading entire directories to xAI’s buckets, the Grok build agent is now public code, enabling audit and fixes.

    xai-org/grok-build, now open source — <p><strong><a href="https://github.com/xai-org/grok-build">xai-org/grok-build, now open source</a></strong></p> xAI's <code>grok</code> CLI tool faced severe community backlash yesterday when it became apparent that running the command in a directory could upload that <em>entire directory</em> to xAI's Google Cloud buckets. One user <a href="https://x.com/a_green_being/status/2076598897779020159">reported</a> running it in their home directory and seeing ...

    signal 8hype 1oss_releasecli_toolsecuritytechnicalsource ↗
  • simonw/blog· AnalysisJul 16, 12:33 AM

    Mermaid to Unicode diagrams

    Simon Willison’s grok-mermaid converts Mermaid graphs into Unicode box art, handy for terminals, diffs, and chat, with results that travel well in text.

    Mermaid to Unicode box art (grok-mermaid) — <p><strong>Tool:</strong> <a href="https://tools.simonwillison.net/grok-mermaid">Mermaid to Unicode box art (grok-mermaid)</a></p> <p>While <a href="https://simonwillison.net/2026/Jul/15/grok-build/">exploring the codebase</a> for the newly open-sourced Grok CLI coding agent I came across <a href="https://github.com/xai-org/grok-build/blob/b189869b7755d2b482969acf6c92da3ecfeffd36/crates/codegen/xai-grok-markdown/src/mermaid.rs">xai-grok-markdown/src...

    signal 7hype 1toolmermaidunicode_arthacksource ↗
  • anthropics/claude-code· First-partyJul 15, 11:02 PM

    Claude Code adds subagent streaming

    v2.1.211 exposes subagent text/thinking in stream-json via a new flag, and hardens approvals against spoofing characters, plus safer PreToolUse handling.

    v2.1.211 — What's changed Added --forward-subagent-text flag and CLAUDE_CODE_FORWARD_SUBAGENT_TEXT environment variable to include subagent text and thinking in stream-json output Fixed permission previews relayed to chat channels not neutralizing bidirectional-override, zero-width, and look-alike quote characters, so tool inputs cannot visually alter the approval message Fixed auto mode overriding a PreToolUse hook's ask decision for unsandboxed Bash — a hook ask now floors the decision at a...

    signal 9hype 1release_notesclaude_codesecurity_hardeningtechnicalsource ↗
  • vercel/next.js· First-partyJul 15, 11:59 PM

    Next.js canary boosts Turbopack cache

    v16.3.0-canary.87 generates component chunks per merged group to raise cache hits, tweaks app shell staleness, and expands docs and bench realism.

    v16.3.0-canary.87 — Misc Changes [turbopack] Generate component chunks for each merged group to increase cache hits: #95261 Exclude stale under 5 minutes from app shells: #95833 docs: remove experimental @next/routing note: #94903 [turbopack] Tell agents not to mention next.config.js options: #95825 chore: Remove stale build warning: #95813 Replay same-document traversals that happen before hydration: #95682 docs: cssChunking graph option: #95693 [Bench] Extend bench app to have realistic cli...

    signal 8hype 1nextjsrelease_notesturbopacktechnicalsource ↗
  • techmeme· AggregatorJul 15, 07:35 PM

    OpenAI unveils GPT-Red red-teamers

    An internal automated model hunts prompt-injection and safety bugs at scale before release, formalizing continuous red-teaming in OpenAI’s deployment pipeline.

    OpenAI details GPT-Red, an internal automated red-teaming model that scales prompt injection vulnerability discovery so it can fix bugs before wider deployment (OpenAI) — OpenAI: OpenAI details GPT-Red, an internal automated red-teaming model that scales prompt injection vulnerability discovery so it can fix bugs before wider deployment — Training strong automated safety red-teamers to improve robustness. — Summary — Problem

    signal 8hype 1model_announcementprompt_injectionred_teaminglaunchsource ↗
  • techmeme· AggregatorJul 15, 11:25 PM

    Cadence launches AuraStack design agent

    New AI platform targets PCB and advanced packaging workflows; early users like Nvidia and TSMC underscore agents moving into high-stakes EDA pipelines.

    Cadence unveils AuraStack AI Super Agent, an AI platform for PCB and advanced chip packaging design, with Nvidia, TSMC, and Schneider Electric among early users (Marco Chiappetta/Forbes) — Marco Chiappetta / Forbes: Cadence unveils AuraStack AI Super Agent, an AI platform for PCB and advanced chip packaging design, with Nvidia, TSMC, and Schneider Electric among early users — As systems and AI infrastructure get more complex, the engineering challenges required to design and bring them to mar...

    signal 6hype 2agent_platformedapcb_designlaunchsource ↗

findings

(04)
  • google/research· First-partyJul 15, 06:06 PM

    Probing diffusion models’ creativity

    Google Research analyzes where “creativity” in diffusion emerges and offers theory-backed metrics, informing controllability and evals for generative image systems.

    Towards demystifying the creativity of diffusion models — Algorithms & Theory

    signal 6hype 1paperdiffusion_modelstheorytechnicalsource ↗
  • huggingface/blog· First-partyJul 15, 05:29 PM

    Lessons from shipping agentic systems

    Hugging Face and AllenAI distill Shippy’s buildout into concrete patterns for tools, memory, and eval loops—useful scaffolding for teams productizing agents.

    What building Shippy taught us about building agents

    signal 6hype 1agentspostmortemcase_studytechnicalsource ↗
  • huggingface/blog· First-partyJul 15, 05:27 PM

    The hard parts of model routing

    IBM Research outlines real-world failure modes and tradeoffs in multi-model routing—accuracy, cost, latency, and drift—plus heuristics to keep routes reliable.

    Model Routing Is Simple. Until It Isn’t.

    signal 5hype 2model_routingllm_opsinferencetechnicalsource ↗
  • hn/frontpage· AggregatorJul 15, 11:03 PM

    Handling NULs in SQLite strings

    SQLite documents how embedded zero bytes behave in TEXT—important for bindings, validation, and security when transporting binary-ish data through SQLite.

    Nul Characters in Strings in SQLite — Article URL: https://sqlite.org/nulinstr.html Comments URL: https://news.ycombinator.com/item?id=48928343 Points: 33 # Comments: 9

    signal 6hype 0sqlitedatabase_internalsstring_encodingtechnicalsource ↗

voices

(01)
  • openai/blog· First-partyJul 15, 12:00 PM

    OpenAI backs reverse-federalism for AI

    OpenAI argues state-led rules can seed a national governance framework—signal on the policy path likely to shape safety standards for AI deployments.

    The US is advancing AI safety through state and federal action — OpenAI outlines a “reverse federalism” approach to AI governance, where state laws help build a national framework for safe, democratic AI.

    signal 5hype 2policygovernanceai_safetyculturalsource ↗