Get tomorrow's brief in your inbox
Today: OpenAI acquired security startup Promptfoo to defend AI agents against attacks. Anthropic launched Code Review to catch bugs in AI-generated code. Nvidia is building an open-source agent platform called NemoClaw for enterprise, while Simon Willison finds coding agents work better with niche tools than expected.
OpenAI Buys Promptfoo: Security for the Agent Era
OpenAI acquired Promptfoo, an AI security startup that helps companies test AI systems for vulnerabilities. The technology will integrate into OpenAI Frontier, the enterprise platform for AI agents. This isn't random shopping. As AI agents get connected to real business data and systems, they become attack surfaces. Bad actors can manipulate prompts, extract sensitive information, or hijack automated workflows. Promptfoo's tools perform automated red-teaming, evaluate agent security, and monitor for compliance risks.
The deal signals that frontier labs know agent security is a make-or-break issue for enterprise adoption. If your AI assistant can read your customer database or execute API calls, you need to know it won't leak data or follow malicious instructions. Promptfoo is used by over 25% of Fortune 500 companies and has raised $23 million since 2024. OpenAI will continue building the open-source version, which lets developers test prompts and compare LLMs like ChatGPT, Claude, and Gemini.
For small businesses: if you're deploying AI agents that touch sensitive data, tools like Promptfoo's open-source library should be part of your testing process. The threat is real, the fix is available, and enterprise customers will expect it.
OpenAI announcement | TechCrunch coverage | CNBC report
Anthropic Code Review catches bugs before merge
Anthropic launched Code Review for Claude Code, an AI reviewer that checks pull requests for logical errors before code ships. It's designed for enterprises drowning in AI-generated code. Claude Code has quadrupled enterprise subscriptions this year and hit $2.5 billion in run-rate revenue, but the flood of pull requests created a review bottleneck. Code Review integrates with GitHub, analyzes every pull request automatically, and leaves comments explaining bugs and fixes. It focuses on logic errors (not style nitpicks) and labels severity with colors: red for critical, yellow for potential issues, purple for preexisting code problems. Multiple agents work in parallel to speed up reviews. Available now in research preview for Claude for Teams and Enterprise customers. If you run a dev team using AI coding tools, this will save your senior engineers from drowning in review queues.
Nvidia preps NemoClaw, an open-source agent platform for enterprises
Nvidia is pitching NemoClaw, an open-source AI agent platform, to enterprise software companies including Salesforce, Cisco, Google, Adobe, and CrowdStrike. The platform will let companies deploy AI agents to perform tasks for employees, with built-in security and privacy tools. It will run on any hardware, not just Nvidia chips. Partnerships would give early access in exchange for contributing to the project. Nvidia CEO Jensen Huang called similar tools like OpenClaw "the most important software release probably ever." The company is betting big on agents as the next major AI category. Formal partnerships aren't confirmed yet, but Nvidia's GPU conference is next week. For small businesses, this signals that enterprise-grade agent tools will soon be available open-source instead of locked behind vendor platforms.
n8n shows how to build multi-domain RAG systems
n8n published a detailed guide on building RAG systems with multiple specialized knowledge bases. The use case: vacation rental managers need separate knowledge bases per property so guests get the right Wi-Fi password or thermostat instructions. The pattern applies to franchise locations, agency clients, or customer support tiers. The workflow uses Google Drive for source files, Pinecone Assistant for semantic search (meaning-based, not keyword-based), and an LLM to generate responses. Pinecone Assistant handles chunking, embeddings, query planning, and result ranking automatically. This is practical for any business managing multiple clients or locations where context matters. If you're building a chatbot or assistant, separating knowledge bases by domain prevents cross-contamination and wrong answers.
IBM releases Granite 4.0 1B Speech: tiny, multilingual, edge-ready
IBM shipped Granite 4.0 1B Speech, a compact speech recognition model for edge devices. It covers English, French, German, Spanish, Portuguese, and Japanese. New features include Japanese ASR support and keyword list biasing (better recognition of names and acronyms). It's half the size of the previous version but delivers higher accuracy and faster inference through speculative decoding. Ranked #1 on the OpenASR leaderboard. Released under Apache 2.0 license with support in transformers and vLLM. If you need on-device transcription without sending audio to the cloud, this is a strong open-source option. Pair it with Granite Guardian for production deployments requiring risk detection.
Coding agents work fine with new tools, not just popular ones
Simon Willison tested whether AI coding agents push developers toward mainstream tools with heavy training data representation. His finding: they don't. He's successfully using agents with brand new tools by feeding them help text and documentation upfront. Modern context windows are long enough to consume substantial docs before starting work. Agents consult existing code examples to learn patterns, then iterate and test output to fill gaps. Willison's testing niche tools like uvx, showboat, rodney, and chartroom. The implication: you can use AI agents with proprietary or custom internal tools by providing good documentation. This flips the "Choose Boring Technology" concern on its head. Skills (specialized agent plugins) are emerging across projects like Remotion, Supabase, Vercel, and Prisma to help agents use them. For small businesses building custom workflows, this means AI tooling won't lock you into mainstream stacks.
Yann LeCun's AMI Labs raises $1.03B to build world models
AMI Labs, cofounded by Turing Prize winner Yann LeCun after leaving Meta, raised $1.03 billion at a $3.5 billion pre-money valuation. The company is building world models: AI that learns from reality, not just language. CEO Alexandre LeBrun predicts "world models" will become the next AI buzzword within six months. First partner is Nabla, a digital health startup. This is fundamental research, not a product sprint. LeBrun says it could take years for world models to reach commercial applications. Co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. Team includes Meta's VP for Europe Laurent Solly as COO and high-profile researchers. For small businesses, this is a signal of where post-LLM AI is headed, but don't expect tools anytime soon.
Nscale hits $14.6B valuation with $2B Series C
British AI infrastructure company Nscale raised $2 billion at a $14.6 billion valuation, making it one of Europe's latest decacorns. The company vertically integrates energy, data centers, compute, and orchestration software. CEO Josh Payne told The New York Times the company might go public "as early as this year." Nscale's "Stargate Norway" project aims to run 100,000 Nvidia GPUs by the end of 2026, with OpenAI as an initial customer. Former Meta COO Sheryl Sandberg, former Yahoo president Susan Decker, and former UK deputy prime minister Nick Clegg joined the board. Backed by Goldman Sachs and JPMorgan. This is pure infrastructure play, not a tool for SMBs, but it shows the compute arms race accelerating.
OpenAI and Google employees support Anthropic's Pentagon lawsuit
Nearly 40 employees from OpenAI and Google, including Google DeepMind chief scientist Jeff Dean, filed an amicus brief supporting Anthropic's lawsuit against the Department of Defense. Anthropic was labeled a supply chain risk after refusing to allow its AI to be used for domestic mass surveillance or fully autonomous weapons. The employees argue the designation is improper retaliation that harms the public interest and that Anthropic's red lines address real risks. The brief states mass domestic surveillance powered by AI poses profound risks to democratic governance and fully autonomous weapons present unaddressed risks. This is a fracture in the AI industry over acceptable use cases for military AI. For businesses evaluating AI vendors, this signals that ethical stances on AI use are becoming differentiators.
The Verge report | TechCrunch coverage
Meta's deepfake moderation called insufficient by Oversight Board
Meta's Oversight Board says Meta's methods for identifying deepfakes are "not robust or comprehensive enough" to handle misinformation during conflicts like the Iran war. The board is calling for Meta to overhaul how it labels AI-generated content across Facebook, Instagram, and Threads. Current systems rely too heavily on self-disclosure and escalated review. Recommendations include improving AI detection tools, scaling "High-Risk AI" labels, and adopting C2PA (Content Credentials) more consistently. The board is concerned Meta is "inconsistently implementing" C2PA even on content from its own AI tools. For businesses using social platforms for marketing or customer communication, expect stricter AI content labeling requirements soon.
X adds Grok photo edit blocker, but it barely works
X introduced a toggle to "block modifications by Grok" when uploading images, but it only prevents tagging @Grok in replies. It doesn't stop users from holding down on images to open "Edit image with Grok" or from saving and re-uploading photos. The feature isn't on web and doesn't apply to older uploads. Grok's image editing was abused to create non-consensual deepfakes in January. Free users are already blocked from editing via @Grok tags, but Premium subscribers can still edit. This toggle is security theater. For businesses posting images on X, assume uploaded photos can be edited by anyone with a Premium account.
Use async RL training to keep GPUs busy
Hugging Face published a deep dive on asynchronous reinforcement learning training. The problem: in synchronous RL, data generation dominates wall-clock time. A single batch of 32K-token rollouts on a 32B parameter model can take hours while training GPUs sit idle. The solution: disaggregate inference and training onto different GPU pools, connect them with a rollout buffer, and transfer weights asynchronously so neither side waits. The survey covers 16 open-source libraries. Key finding: Ray dominates orchestration. If you're training models post-ChatGPT or running RL workflows, this pattern keeps GPUs from idling 60% of the time. Read the full comparison table for implementation details.
OpenAI's Promptfoo acquisition and Anthropic's Code Review launch signal that AI security and code quality are front-burner issues, not future concerns. As agents move from demos to production systems touching real data, the industry is scrambling to build guardrails. Nvidia's NemoClaw pitch shows enterprise agent platforms are emerging open-source, not just from API vendors. And the Anthropic-Pentagon standoff, backed by employees across AI labs, marks the first major industry fracture over acceptable AI use cases. For businesses deploying AI tools this year, expect security tooling, code review automation, and vendor ethical stances to become selection criteria.