“xAI released the Grok Build codebase under an Apache 2.0 license after backlash over a feature that could upload entire directories; Willison inspects the large Rust codebase, highlights interesting files (including a Mermaid renderer), and notes privacy-related changes by xAI.”

#16 📝 Simon Willison xai-org/grok-build, now open source - xAI released the Grok Build codebase under an Apache 2.0 license after backlash over a feature that could upload entire directories; Willison inspects the large Rust codebase, highlights interesting files (including a Mermaid renderer), and notes privacy-related changes by xAI. He also got a version of the Mermaid renderer working in WebAssembly. #17 📝 Simon Willison How I tricked Claude into leaking your deepest, darkest secrets - Ayush Paul demonstrated a web_fetch exfiltration loophole in Claude by chaining nested links inside fetched pages, enabling data exfiltration from the model's memory; Anthropic has closed the hole by preventing web_fetch from navigating to links embedded in fetched content. The attack extracted user's name, home city and employer in the demonstration.

“xAI has installed Grok Build into Railway sandboxes, letting PM Builders instantly spin up and test AI-powered workflows in their dev environment.”

GenAI PM Daily July 03, 2026 GenAI PM Daily 🎧 Listen to this brief 3 min listen Today's top 23 insights for PM Builders, ranked by relevance from Blogs, X, and YouTube. DeepAgents introduces dynamic subagents for modular AI workflows #1 📝 Anthropic News More details on Fable 5’s cyber safeguards and our jailbreak framework - Claude Fable 5 has been re‑deployed and is available globally, and Anthropic trained safety classifiers to sort cybersecurity requests into four categories—Prohibited (blocked), High‑risk dual use (blocked), Low‑risk dual use (monitored/sometimes blocked within a larger “safety margin”), and Benign (allowed with monitoring)—with Fable 5’s safety margin larger than prior models. Anthropic also published an early draft jailbreak severity framework with Glasswing, launched a HackerOne program and feedback channel (cyber-safeguards@anthropic.com) for reporting cyber jailbreaks, and enumerated prohibited actions such as ransomware, cyber‑physical sabotage, AV/EDR bypass, command‑and‑control, data exfiltration, malware development/propagation, and internet backbone attacks. Also covered by: @Claude #2 📝 Claude Code Blog Giving admins more visibility and control over Claude spend - Announces new admin features that give organizations greater visibility into Claude usage and spending, along with controls to manage costs and governance. The update targets product administrators who need reporting and spend management tools. #3 𝕏 xAI has installed Grok Build into Railway sandboxes, letting PM Builders instantly spin up and test AI-powered workflows in their dev environment. #4 𝕏 Thariq confirms Fable will exit subscription plans on July 7 but will be reinstated as a standard feature as soon as capacity allows, per the original blog post. Also covered by: @Claude #5 𝕏 Harrison Chase introduced dynamic programmatic subagents in DeepAgents, letting developers spin up on-the-fly subagents—each with its own memory, tools, and prompts—to modularize and scale complex AI workflows. #6 𝕏 Santiago demoed an agent built on the Linux Foundation-governed x402 open protocol (by Coinbase) that autonomously discovers, pays (in USDC on Base via HTTP 402), and runs Apify Store Actors with no API keys, accounts, or manual billing—enabling true pay-as-you-go agent workf... #7 𝕏 LlamaIndex 🦙 built an email-processing assistant using LiteParse inside a flueai agent, with Resend webhooks and tursodatabase for persistence, that fetches attachments (including PDFs), parses them, summarizes messages, and drafts replies. #8 𝕏 Philipp Schmid built Gemini Omni Flash in just 12 lines using the Interactions API to let you conversationally edit video lighting—upload a clip, request “Make it day time,” and get back a video with shifted shadows and a new sky. #9 𝕏 Guillermo Rauch announces AI Gateway Rules—a CDN-style feature to dynamically reroute or deny AI model traffic on the fly (e.g., rewrite anthropic/claude-fable-5 → anthropic/claude-opus-5)—so you can handle sudden model retirements without redeployment. #10 📝 Ampcode Chronicle Putting an Agent in an Orb - Amp runs orbs on Debian 12 with preinstalled dev tooling (gh, amp, PostgreSQL, Redis, tmux, ffmpeg, ImageMagick, ripgrep, Bun/Node, npm/pnpm, agent-browser), and a 428-line .agents/setup script that starts PostgreSQL (tuned for speed: fsync=off, synchronous_commit=off, autovacuum=off), creates the amp user/database, seeds test users, installs mise and repo-managed Node/pnpm, runs pnpm install --frozen-lockfile, installs Pillow and sqlite3, writes AGENTS.md guidance, and snapshots the orb for up to 24 hours. The repo also provides a dev-server skill with ensure-dev-server.sh (which ensures .env.local and secrets AMP_API_KEY_SECRET, THREAD_ACTORS_SERVICE_SECRET, THREAD_ACTORS_WEBSOCKET_JWT_SECRET, and can reuse/restart/start the server), writes .amp/dev-ports.json, and exposes /__dev endpoints (including /__dev/log-me-in/ , /__dev/sudo, /__dev/preflight) so agents can autonomously start, authenticate to, and verify the dev environment and produce screenshots/results in minutes. #11 📝 Ampcode Chronicle Read Bigger Threads - Amp rewrote read_thread into a dedicated subagent (now running GLM 5.2 instead of Gemini 3.5 Flash) that searches long threads, verifies whether edits actually succeeded, and explicitly checks for newer messages that revise, supersede, or revert earlier hits. The change was driven by compaction making threads huge—one thread would be ~21 million tokens without compaction and has been compacted over 68 times—and the subagent uses compactions for orientation but inspects original messages when exact wording, code, chronology, or verification matter. #12 📝 Simon Willison llm-coding-agent 0.1a0 - Simon shipped an early alpha of a new Python coding agent built on his LLM/agent framework, created via Claude Fable 5 and released to PyPI and GitHub. The post describes the prompts, spec, README, available tools, and a demo transcript showing the agent generating a simple Swift CLI ASCII clock. #13 📝 Simon Willison Using DSPy to evaluate and improve Datasette Agent's SQL system prompts - Simon experimented with dspy to evaluate and improve the system prompts used by Datasette Agent for read-only SQL queries, firing off an asynchronous research task using Claude Fable 5. The tool suggested testing with smaller models and highlighted prompt issues like missing column names in schema listings, which can cause guessing and retry loops. #14 ▶️ 160,000+ Cloned These 3 FREE AI Employees: Here's How (GitHub Claude Skills) Helena Liu Installing three free GitHub AI agent repositories—LLM Console skill, Last30Days, and G-Stack—into Claude CoWork desktop via clone commands to enable five debating advisors, comprehensive 30-day sentiment analysis, and a Y Combinator–style AI development team through slash commands. LLM Console skill (based on Karpathy’s method) installs five advisors—Contrarian, First Principle Thinker, Expansionist, Outsider, Executor—via git clone into Claude CoWork and is activated with “console this” for strategic business debates. The Last30Days repository, with over 46,000 GitHub stars, uses the slash command “/last30days <topic>” to scan Reddit, X, YouTube, TikTok, Instagram, and HackerNews and returns summarized sentiment with exact URLs in about five minutes. G-Stack by Garry Tan (115,000+ GitHub stars) installs via a single clone command into Claude CoWork to provide roles like CEO, engineering manager, senior designer, QA lead, and tester, accessible through dedicated slash commands such as “/officehour”. #15 𝕏 Boris Cherny uses Claude to auto-generate artifacts—tables, diagrams, color-coded charts and PR overviews—to streamline architecture/design option comparisons, session result visualization, data analysis and complex code reviews. #16 𝕏 Lenny Rachitsky relays OpenAI’s Codex app lead caution that eliminating dedicated product roles to make “everyone a builder” is a mistake, because product management is a distinct discipline with unique best practices and skills engineers often overlook. #17 𝕏 Aravind Srinivas argues that personal robots are the killer app for running AI models on local hardware—people won’t stream home data to servers, and such devices can double as token faucets for digital tasks. #18 𝕏 Lenny Rachitsky asked OpenAI Codex lead @ajambrosino why AI “sucks at design.” He says labs prioritize coding—because it’s easier to grade and accelerates AI research—while real design demands cultural insight and novelty, so models default to familiar patterns (e.g. #19 𝕏 Julien Chaumond – Co-founder and CTO @huggingface notes that even Palantir now publicly supports open models, underscoring a significant industry shift toward AI openness. #20 𝕏 clem 🤗 – Co-founder & CEO @HuggingFace Rampart from @ndstudio & the @WhiteHouse is now the #1 trending token classification model on Hugging Face, marking a shift as public organizations build and own their own model weights instead of renting them. #21 ▶️ Claude Fable 5 Is Finally Back: 5 Must-Try Use Cases Before July 7 Peter Yang Demonstrates five specific high-leverage use cases for Claude Fable 5 on a Claude subscription—via the Claude Code interface with high-effort prompts and API integrations—before its July 7 availability cutoff at 50% weekly usage. Claude Fable 5 is available on Claude subscriptions until July 7 with a 50% weekly usage cap, after which it switches to pay-as-you-go API credits. Fable 5 audited a vibe-coded fitness app via five parallel agents, running all unit tests and identifying over 12 major bugs including a user-data leak on involuntary sign-out. Fable 5 drafted a detailed HTML plan for a new nutrition-tracking tab—complete with UI mockups, Supabase schema, and USDA FDC API integration—using the Lavish editor plugin. #22 ▶️ Fable 5 vs GPT 5.6 Sol: The Early Results AI Explained GPT 5.6 Soul is benchmarked against Anthropic’s Fable 5 (Mythos 5) on Terminal Bench 2.1, Healthbench Professional, and Exploit Bench, with GPT 5.6 Soul scoring almost 92% vs 88% on Terminal Bench, 60.5% (64% length-adjusted) vs 66.0% on Healthbench, ≈76% vs ≈78% on Exploit Bench, and costing half the API input price and just over half the output price of Fable 5. GPT 5.6 Soul on Terminal Bench 2.1 ultra mode scored almost 92% compared to Mythos 5’s 88%. On Healthbench Professional, Mythos 5 scored 66.0% raw horsepower versus GPT 5.6 Soul’s 60.5% (64% length-adjusted). GPT 5.6 Soul used approximately 120,000–130,000 output tokens on Exploit Bench, compared to 350,000 tokens for Mythos 5. #23 𝕏 Harrison Chase unveils LangSmith Engine, an agent launched last month that hunts through your AI agents’ failures, prioritizes issues, and auto-drafts fixes using sandboxed environments and subagents.

“xAI launched Voice Agent Builder, a no-code platform to create human-like voice agents using Grok Voice, available today at $0.05/min.”

#1 𝕏 xAI launched Voice Agent Builder, a no-code platform to create human-like voice agents using Grok Voice, available today at $0.05/min. #25 𝕏 xAI launched a turnkey production voice agent platform featuring telephony, knowledge retrieval, developer tools, guardrails, and observability.

“#11 𝕏 xAI integrated its SpaceXAI state-of-the-art voice APIs into the Vercel AI Gateway, giving developers seamless access to high-fidelity speech synthesis and recognition tools.”

#11 𝕏 xAI integrated its SpaceXAI state-of-the-art voice APIs into the Vercel AI Gateway, giving developers seamless access to high-fidelity speech synthesis and recognition tools.

“xAI integrated SuperGrok and X subscriptions into T3code, letting developers access their AI-powered features directly within the coding environment.”

#15 𝕏 xAI integrated SuperGrok and X subscriptions into T3code, letting developers access their AI-powered features directly within the coding environment.

“xAI launched the official @MongoDB plugin for Grok Build, enabling users to query data, optimize indexes, and manage databases directly within the tool.”

xAI appears as the company behind Grok Build and the MongoDB plugin announcement. The emphasis is on database operations inside an AI builder environment.

“xAI now integrates Grok with Interactive Brokers, delivering high-quality, real-time portfolio information.”

This item notes an integration between xAI’s Grok and Interactive Brokers for real-time portfolio information.

“xAI added its Grok models to Databricks Agent Bricks, letting enterprises bring SpaceXAI’s latest AI to their data for powering capable AI agents.”

📝 𝕏 xAI added its Grok models to Databricks Agent Bricks, letting enterprises bring SpaceXAI’s latest AI to their data for powering capable AI agents.

“xAI launched Grok 4.3 on Amazon Bedrock, giving AWS developers access to the industry-leading low-hallucination model with advanced tool calling via Bedrock’s secure inference engine.”

#7 𝕏 xAI launched Grok 4.3 on Amazon Bedrock, giving AWS developers access to the industry-leading low-hallucination model with advanced tool calling via Bedrock’s secure inference engine. #8 𝕏 xAI launched Imagine Video 1.5 in its API and rolled out Video 1.5 Fast for consumers, cutting 720p render times from 40+ seconds to about 25 seconds while boosting quality.

“xAI now lets you tap your SuperGrok or X Premium subscription directly in Warp (warpdotdev)—just head to Warp Agent Settings and switch to the Grok Build model.”

#2 𝕏 xAI now lets you tap your SuperGrok or X Premium subscription directly in Warp (warpdotdev)—just head to Warp Agent Settings and switch to the Grok Build model.