Today’s lead · techcrunch.com
U.S. lifts Anthropic Fable/Mythos export controls; Claude Fable 5 returns with safeguards
The U.S. dropped the license requirement that had forced Anthropic to pull public access to Mythos and Fable, and Anthropic said it would begin restoring access on July 1. The company agreed to detect and address security risks, work with the U.S. government on protocols and standards for current and future releases, and report malicious activity.
Top signals
4 more
Cloudflare sets a September 15 deadline for AI crawlers to split search from training and agent use
Cloudflare says that starting September 15, 2026, its default settings will block mixed-use crawlers from ad-supported pages for new Cloudflare customers, new sites from existing customers, and all existing free customers unless site owners opt otherwise. It is also evolving Pay Per Crawl into Pay Per Use with initial partners Ceramic.ai and You.com.
Cursor 3.0 patches DuneSlide zero-click sandbox escapes
Cato AI Labs disclosed two Cursor flaws, CVE-2026-50548 and CVE-2026-50549, that could let a poisoned prompt delivered through MCP-connected services or web search escape the editor sandbox and run commands on a developer machine. Versions before Cursor 3.0 are affected; the fixes shipped in Cursor 3.0.
Meta is reportedly preparing an AI compute and model cloud business
Bloomberg reports that Meta is developing plans to sell access to AI compute power and models through a cloud infrastructure business, with TechCrunch noting the initiative is reportedly called Meta Compute. Meta has committed $182.9 billion to AI infrastructure in coming years and is looking for returns beyond its own AI products.
Unpatched Argo CD repo-server flaw can expose Kubernetes clusters to takeover paths
Synacktiv disclosed an unpatched Argo CD repo-server flaw that can let an unauthenticated attacker who can reach the internal gRPC port execute code through Kustomize’s helm-command option. The report says there is no fix or CVE yet, and recommends enabling Kubernetes network policies around repo-server and Redis.
Tools & repos
6 selected
Tabstack Browser Automation
Tabstack offers hosted browser automation for apps and agents, so teams can drive web workflows without operating their own browser fleet. It is most relevant for agent builders that need reliable web actions as an infrastructure primitive.
Exa MCP
Exa’s MCP now exposes web search, code search, Exa Agent, and Exa Connect inside Claude, Codex, Cursor, VS Code, and other MCP clients. It gives coding and research agents a configurable search and enrichment layer without each team wiring those tools from scratch.
usestrix/strix
Strix is an open-source AI penetration testing tool for finding and fixing application vulnerabilities. It is worth evaluating as teams move security testing earlier into agent-assisted development and pre-deploy checks.
xAI Voice Agent Builder
xAI launched a no-code Voice Agent Builder for creating human-like agents with Grok Voice, priced at $0.05 per minute. The launch positions voice agents as an operations tool for support, sales, scheduling, handoffs, and phone workflows rather than a standalone voice demo.
Gemini Spark for macOS
Google’s Gemini Spark agent is now in the Gemini Mac app beta for U.S. Google AI Ultra subscribers, with local-file workflows plus integrations for Tasks, Keep, Canva, Dropbox, Instacart, OpenTable, and Zillow Rentals. Google also says custom MCP support is rolling out, making Spark more relevant as a desktop agent surface.
facebook/astryx
Astryx is Meta’s open-source, customizable design system described as agent-ready. For product teams using coding agents to build interfaces, a structured design system can reduce inconsistent AI-generated UI work.
Blogs worth your time
6 reads
How to Use RLMs in Deep Agents
LangChain explains how Deep Agents now uses dynamic subagents and a lightweight code interpreter to run recursive, code-driven workflows over large context instead of relying on one turn-by-turn agent loop. The useful pattern: let the model write map/reduce-style orchestration code, then dispatch subagents for deterministic coverage across huge inputs.
OpenWiki: Open Source Repo Documentation for Coding Agents
OpenWiki is an open-source agent and CLI that generates a repo wiki, points AGENTS.md or CLAUDE.md at it, and can keep it current through a GitHub Action. It is a practical answer to a common coding-agent failure mode: dumping too much stale repo context into one instruction file.
The bottleneck moved to verification. So we automated that too.
Augment describes an internal Verifier agent that deploys each PR to an isolated live environment, exercises the changed behavior, and posts logs, screenshots, API responses, and limits back to reviewers. The post is a useful blueprint for moving agentic coding from “generated code” to runtime evidence.
Mastering Agentic Techniques: AI Agent Reinforcement Learning
NVIDIA lays out when to use prompting, SFT, DPO, RLHF, or RLVR for agent behavior, then walks through building a small GRPO loop with verifiable rewards. The practical takeaway is to start with evals and trusted verifiers before training; otherwise RL just optimizes the wrong shortcut faster.
Multi-token Residual Prediction
Modal and NYU Shanghai show how Multi-token Residual Prediction adds a small frozen-backbone module to diffusion language models to predict inter-step residuals rather than full next-step distributions. The result is a concrete inference knob: lossless speculative speedups or faster direct decoding with measurable quality tradeoffs, with code and an SGLang implementation available.
Hugging Face and Cerebras bring Gemma 4 to real-time voice AI
Hugging Face details an open, cascaded speech-to-speech stack using Parakeet for ASR, Gemma 4 31B on Cerebras for inference, and Qwen3TTS for speech output. It is useful for teams building voice assistants or robots because it decomposes the latency problem into replaceable open components instead of a black-box voice API.
Funding & acquisitions
3 moves
Venice AI
Venice AI raised a $65 million Series A at a $1 billion valuation, led by Dragonfly with participation from Coinbase Ventures, North Island Ventures, and others. The privacy-focused AI platform says it is profitable, has over $70 million in annualized run-rate revenue, more than 3 million active users, and averages 1.7 million API calls per day.
Wayve
Wayve launched an $85 million employee tender offer at an $8.5 billion valuation, led by existing and new investors. The autonomous-driving startup has grown to 1,200 employees, is targeting robotaxi pilots with Uber later this year, and plans to integrate its AI software into Nissan driver-assist systems starting in 2027.
HIVE Digital Technologies
HIVE Digital Technologies closed a US$130 million 0% exchangeable senior notes offering, bringing its total 0% coupon financing this quarter to US$245 million. The company says the capital will accelerate GPU acquisitions and expansion of its AI and HPC infrastructure platform.
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.