OpenAI Reveals Jalapeño, Its First Custom AI Chip The Verge / TechCrunch — OpenAI OpenAI and Broadcom unveiled Jalapeño, an ASIC “Intelligence Processor” purpose-built for LLM inference, taken from first design to tape-out in just nine months — a record for an advanced chip of its kind. The company used its own models to accelerate parts of the chip design process, and positions Jalapeño as a step toward full hardware independence from Nvidia. Jalapeño will power ChatGPT, Codex, the API, and future agentic products.

🔒 Security & Policy

NSA Lost Access to Powerful AI Model Amid Anthropic Dispute — New York Times Anthropic’s Mythos model reportedly identified vulnerabilities in nearly all NSA classified systems within hours during an authorized red-team evaluation — a result that stunned analysts and is now reported to be the central trigger for the US government imposing export controls directly on an AI model for the first time. The controls forced Anthropic to abruptly disable Fable 5 and Mythos 5 for all customers globally, including Five Eyes government partners who received no advance warning. Anthropic is working to restore access under a compliance framework being negotiated with the White House.

Insights on Indirect Prompt Injection — Latent Space Gray Swan’s founders discuss the accelerating focus on jailbreaks and indirect prompt injection attacks, the company’s role evaluating frontier AI systems, and how they build security benchmarks for models that increasingly have access to real-world tools.

Prompt Injection as Role Confusion — role-confusion.github.io Researchers argue prompt injection is a structural flaw: LLMs receive all input through the same token stream and cannot reliably distinguish operator instructions from adversarial content injected in data, making injection defense a perpetual whack-a-mole problem.

US Presses Meta to Agree to AI Reviews as Security Concerns Rise — New York Times The Trump administration is pushing Meta — the only major US AI developer without a voluntary model-sharing agreement — to submit its models for federal capability and vulnerability evaluations. Meta’s policy team is negotiating with the Commerce Department but no agreement has been reached.

Patch the Planet: OpenAI + Trail of Bits Target Open-Source Vulnerabilities — OpenAI OpenAI’s Daybreak initiative has partnered with Trail of Bits to identify and fix security flaws in critical open-source projects including cURL, Python, and Go, with hundreds of issues already found and dozens of patches merged using AI-assisted research.

Vulnerability Reports Are Not Special Anymore — words.filippo.io A sharp take arguing that AI has shifted the bottleneck in security from finding vulnerabilities to assessing which reported issues are actually real — a signal for how security triage workflows will need to evolve.

🤖 Frontier Models & Silicon

ByteDance’s Seedance 2.5 Generates 30-Second 4K Video from a Single Prompt — CNET Seedance 2.5 can produce 30-second 4K video clips from a text prompt, with support for up to 50 reference images, videos, or audio clips to give users fine-grained creative control. The model launches in China next month; no international release window has been announced.

Mistral OCR 4: SOTA Document Intelligence — Mistral AI Mistral’s OCR 4 offers structured content extraction with bounding boxes and confidence scores across 170 languages, ships as a single container, and claims 4× speed advantages over competing systems — particularly strong on low-resource languages.

OpenAI Prepares Bidirectional Voice Mode (Bidi 1) for ChatGPT — Testing Catalog OpenAI’s Bidi 1 lets ChatGPT speak, hear, and listen simultaneously — handling mid-conversation interruptions and real-time translation. Some users are already seeing it in model selectors ahead of a formal announcement; copyright restrictions apply to generative singing.

Krea 2 Technical Report — Krea Krea 2 uses a multi-stage training process with a prompt expander and style-reference system to push image generation toward genuine stylistic diversity and user control, explicitly designed to escape the homogeneous “default aesthetic” trap.

🛠️ Developer Tools & Agents

Anthropic Launches Claude Tag for Slack — Anthropic Claude Tag lets teams @-mention Claude in any Slack channel, assign it async tasks, and connect it to codebases and tools with context retained across conversations. Anthropic says its own product team already generates much of their code this way. Now in beta for Enterprise and Team users, it runs on Opus 4.8 and replaces the existing Claude Slack app.

Figma Unveils Full-Stack Canvas, AI Motion Graphics, and Shader Tools at Config 2026 — TechCrunch Config 2026 brought a reimagined canvas optimized for full-stack development that integrates AI agents, design tools, and code in one environment. New AI-powered motion graphics and shader tools target creative production workflows, and a new code layer tightens the handoff between design and engineering.

AWS Lambda Introduces MicroVMs for Sandboxed Code Execution — AWS Lambda MicroVMs provides VM-level isolation with near-instant startup powered by Firecracker, supporting up to 16 vCPUs, 32 GB RAM, and 8-hour sessions with suspend/resume. Explicitly designed for running untrusted user or AI-generated code safely at scale.

Analyzing Claude Code Usage with CloudWatch and OpenTelemetry — AWS Amazon CloudWatch OTLP now ingests Claude Code telemetry directly via bearer token — no collector or IAM setup required — enabling per-developer cost attribution, PromQL dashboards, and productivity analytics across teams.

Amazon ECS Adds 20-Second High-Resolution Metrics for Faster Autoscaling — AWS

How Netflix Simplified Batch Compute with Kueue — Netflix Tech Blog

Build and Deploy a Remote MCP Server to GKE in 30 Minutes — Google Cloud

The Coming Loop — Armin Ronacher A thoughtful critique of agent loops: while they enable astonishing build speed, they risk reducing developers to messengers who outsource judgment to the machine — and with it, the engineering discipline that makes software trustworthy.

💰 Investment & Industry

Agility Robotics Plans to Go Public via $2.5B SPAC Deal — TechCrunch Agility Robotics, maker of the Digit humanoid, is merging with Churchill Capital XI at a $2.5B valuation, expecting $620M in proceeds from a PIPE backed by NVIDIA, Amazon, SoftBank, and Foxconn. Digit is already deployed commercially with GXO, Toyota, and Schaeffler, making Agility the first pure-play US humanoid robotics company to list publicly.

$27M AI Proxy War Over Congressional Seat Ends in a Draw — The Verge A $27M political proxy battle between Anthropic- and OpenAI-backed super PACs targeting New York’s 12th Congressional District primary ended inconclusively, with AI-aligned candidate Alex Bores narrowly losing the Democratic race.

After Betting the Firm on Anthropic, Menlo Ventures Raises a $3B Fund — TechCrunch Menlo Ventures is closing a $3B Fund VII, the payoff from a single bold $750M Anthropic investment in 2024 that redefined the firm’s identity as an AI-focused investor.

Meta Announces New Smart Glasses Starting at $299 — CNBC

Facebook Rolls Out AI Companion App for Creators — TechCrunch

Generated by claude-sonnet-4-6 · 2026-06-24T10:00:00Z