Claude Code bug silently charges Max plan users extra after HERMES.md string triggers anti-abuse system

A Claude Code user discovered that having the case-sensitive string 'HERMES.md' in git commit messages causes API requests to be routed to extra usage billing instead of the included Max plan quota. This silently consumed $200.98 in extra usage credits while the Max 20x plan capacity remained largely unused. The bug was traced to an overactive anti-abuse system server-side that triggered on this specific string in the system prompt context. Anthropic acknowledged and fixed the bug but initially refused to issue a refund, sparking significant community backlash in the issue thread.

LLM 0.32a0 adds typed message sequences and streaming event parts

LLM 0.32a0 is an alpha release of Simon Willison's LLM Python library and CLI tool, introducing two major backwards-compatible changes. First, prompts can now be expressed as a sequence of typed messages (user/assistant turns) rather than just a single text string, enabling pre-built conversation injection without SQLite dependency. Second, streaming responses now emit typed event parts (text, reasoning, tool call names, tool call args) instead of raw text chunks, allowing consumers to handle mixed-content responses from modern models. The release also adds response serialization/deserialization via to_dict()/from_dict(), and a new -R/--no-reasoning CLI flag. A redesign of the SQLite logging system to support graph-based conversation storage is planned for a future release.

Parallel Web Systems raises $100M Series B at $2B valuation for AI agent search APIs

Parallel Web Systems, the AI agent-tool startup founded by former Twitter CEO Parag Agrawal, has raised a $100 million Series B at a $2 billion valuation led by Sequoia. This comes just five months after its $100 million Series A at a $740 million valuation, bringing total capital raised to $230 million. The company offers web search and research APIs designed specifically for AI agents, with customers including Clay, Harvey, Notion, and OpenDoor, as well as unnamed banks and hedge funds. It reports over 100,000 developers using its products.

Stripe Sessions 2026 launches Machine Payments Protocol and agentic commerce infrastructure

<h2>Stripe Sessions 2026: 288 announcements</h2>
<p>Stripe held its annual Sessions conference and announced 288 new products and features. The breadth is hard to summarize cleanly, but a few threads run through most of it.</p>
<p><strong>Agentic commerce</strong> is the clearest theme. Stripe expanded its Agentic Commerce Suite with partnerships from Meta and Google, enabling AI agents to make purchases autonomously. Alongside that, they introduced the Machine Payments Protocol (MPP), designed specifically for agent-to-agent microtransactions — the idea being that agents need their own payment rails, not just access to human ones.</p>
<p>Other notable announcements:</p>
<ul>
<li><strong>Checkout Studio</strong> with AI-assisted optimization and A/B testing</li>
<li><strong>Radar upgrades</strong> including bot abuse prevention and custom fraud models</li>
<li><strong>Usage-based billing</strong> improvements via Metronome</li>
<li><strong>Stripe Database</strong>, a real-time Postgres offering</li>
<li><strong>Treasury</strong> expanding to Australia and Canada, with stablecoin support</li>
<li><strong>Stripe card</strong> with 2% cashback</li>
<li><strong>Global Payouts</strong> to 100+ countries</li>
<li><strong>Issuing for AI agents</strong> — agents can now be issued cards directly</li>
<li>A <strong>public product roadmap</strong> through Q1 2027</li>
</ul>
<h2>Stripe Projects: let agents provision infrastructure</h2>
<p>The more immediately interesting thing, at least for developers, is Stripe Projects — currently in open beta. It’s a CLI workflow that lets developers (or AI coding agents) provision and configure third-party services from the terminal without touching any dashboards.</p>
<p>The basic flow looks like this:</p>
<pre><code>stripe projects init my-app
stripe projects <span class="hljs-built_in">add</span> auth0<span class="hljs-built_in">/client
</span></code></pre>
<p>That second command provisions Auth0 resources and returns credentials in a format agents can read directly. Auth0 is now available in the developer preview, and the pattern is designed to be extensible — any platform with signed-in users can act as an orchestrator the same way Stripe does.</p>
<h2>Cloudflare integration: agents that deploy themselves</h2>
<p>The most striking demonstration of where this is heading: Cloudflare and Stripe have built a protocol that lets AI coding agents autonomously create Cloudflare accounts, start paid subscriptions, register domains, and deploy applications — with no human in the loop.</p>
<p>The protocol is built on OAuth, OIDC, and payment tokenization. Three components make it work:</p>
<ol>
<li><strong>Service discovery</strong> via a catalog API</li>
<li><strong>Identity-based authorization</strong> with Stripe acting as the identity provider</li>
<li><strong>Payment tokenization</strong> with a default $100/month spending cap</li>
</ol>
<p>Agents use the Stripe Projects CLI to orchestrate the whole flow. Cloudflare’s Code Mode MCP server and Agent Skills are included to improve deployment quality. The $100/month cap is a sensible guardrail — though it’s worth noting that a sufficiently motivated agent could still rack up meaningful costs before anyone notices.</p>
<p>The tweet that circulated around this was blunt about what it means: let your agent sign up for Cloudflare, buy a domain, and deploy to production, all on its own.</p>
<h2>Google Cloud Agents CLI</h2>
<p>Separately, Google Cloud launched Agents CLI as part of its Agent Platform. The goal is similar — reduce the friction of building and deploying AI agents — but the focus is more on the development lifecycle than on autonomous provisioning.</p>
<p>It integrates with coding agents like Gemini CLI, Claude Code, and Cursor, and provides structured access to Cloud Run and Kubernetes. Key features include local simulation and evaluation pipelines, automated infrastructure-as-code and CI/CD generation, and a “Human Mode” that lets developers inspect and control workflows rather than handing everything to the agent.</p>
<p>That last feature is interesting. There’s a real tension right now between agents that are genuinely autonomous and agents that still need a human to catch mistakes. Human Mode is Google’s acknowledgment that most teams aren’t ready to fully let go yet.</p>
<h2>What this adds up to</h2>
<p>Taken together, these announcements describe a world where AI agents don’t just write code — they provision infrastructure, register domains, start subscriptions, and deploy to production. Stripe is positioning itself as the identity and payment layer for that world, which is a significant bet.</p>
<p>Whether agents doing all of this autonomously is exciting or unsettling probably depends on how much you trust the spending caps and authorization flows. The technical pieces are coming together faster than most people expected. The governance pieces are still catching up.</p>


Stripe's Sessions 2026 conference featured 288 announcements centered on agentic commerce, including a Machine Payments Protocol for agent-to-agent microtransactions, Stripe Projects (open beta CLI for provisioning third-party services like Auth0), and a Cloudflare integration enabling AI agents to autonomously create accounts, register domains, and deploy apps with a $100/month spending cap. Google Cloud also launched an Agents CLI integrating with Gemini CLI, Claude Code, and Cursor for streamlined agent development and deployment. Together, these moves position Stripe as the identity and payment layer for a world where AI agents autonomously provision infrastructure and transact.

Stripe launches Link wallet for agents with one-time virtual cards and spend approvals

Stripe is launching Link's wallet for agents, a new product that gives AI agents programmatic access to payment credentials without exposing raw card data. Built on Stripe's new Issuing for agents, it allows agents to request one-time-use virtual cards or Shared Payment Tokens (SPTs) backed by a user's existing Link wallet. Consumers approve each spend request via web or mobile app. For developers building consumer-facing agents, it removes the need to build wallet infrastructure from scratch. Businesses can also use Stripe Issuing for agents directly to build custom agentic wallets, set spending controls, automate business spend, and manage fund flows for use cases like fintech, vertical SaaS, and marketplaces.

Google CEO says 75% of new code is now AI-generated, up from 25% eighteen months ago

Google CEO Sundar Pichai disclosed that 75% of Google's new code is now AI-generated, up from 25% just 18 months ago. Engineers have shifted from writing code to orchestrating autonomous AI agents that plan, execute, and refactor entire codebases. A complex code migration was completed 6x faster than a year prior. Other tech giants are following suit: Meta targets 65% AI-assisted code commits, Snap reports 65% AI-generated code, and Microsoft's CTO predicts 95% AI-generated code within five years. The shift is reshaping the career ladder, with a Stanford analysis showing a 20% drop in employment among junior developers since late 2022. Internally, Google DeepMind teams have reportedly sought to use Anthropic's Claude Code alongside Gemini, creating friction around tool standardization.

Anaconda acquires Outerbounds to bring Metaflow orchestration to AI-native development

Anaconda is acquiring Outerbounds, the company behind Metaflow, the open source AI/ML orchestration framework originally built at Netflix. The deal aims to create a full-stack, governed platform for AI-native development. The acquisition is driven by a growing enterprise problem: AI-generated code now makes up nearly half of new enterprise code, but produces 1.7x more defects than human-written code, and 80% of AI-recommended dependencies carry known security risks. By combining Anaconda's trusted Python dependency stack with Metaflow's production-grade orchestration, the combined platform targets the entire AI development lifecycle — from secure experimentation to production — running on customer-owned infrastructure without cloud lock-in. Anaconda has committed to continuing Metaflow as an open source project.

Mistral releases Medium 3.5 128B open-weights model with remote coding agents and 256k context

<p>Mistral AI released Mistral Medium 3.5, a 128B dense model with a 256k context window and configurable reasoning effort. It scores 77.6% on SWE-Bench Verified and is now the default model in both Vibe CLI and Le Chat.</p>
<h2>What’s new</h2>
<p>The release ships with remote coding agents in Mistral Vibe. These run asynchronously in the cloud, support parallel sessions, and integrate with GitHub, Linear, Jira, Sentry, and Slack. You can launch them from either the CLI or Le Chat.</p>
<p>Le Chat also gets a new Work mode (currently in preview) that handles multi-step agentic tasks - things like research, inbox triage, and workflows that span multiple tools.</p>
<h2>Availability and pricing</h2>
<p>The model is available via API at $1.50 per million input tokens and $7.50 per million output tokens. Weights are on Hugging Face under a modified MIT license.</p>


Mistral AI has released Mistral Medium 3.5, a 128B dense open-weights model featuring a 256k context window and configurable reasoning effort. It achieves 77.6% on SWE-Bench Verified and is now the default model in Vibe CLI and Le Chat. The release introduces remote coding agents in Mistral Vibe that run asynchronously in the cloud, support parallel sessions, and integrate with GitHub, Linear, Jira, Sentry, and Slack. Le Chat gains a new Work mode (preview) for multi-step agentic tasks. The model is available via API at $1.50/$7.50 per million input/output tokens, with weights on Hugging Face under a modified MIT license.

OpenCode 2.0 adds distributed sessions with automatic data replication and server resync

OpenCode 2.0 is introducing improved embeddability, including a workspaces feature for running distributed sessions. Data replication is handled automatically, abstracting away the remote execution complexity. Sessions are resilient to server downtime, with automatic resync when the server comes back online.

Nuxt launches first-party AI agent built on its MCP server with StackBlitz and streaming support

Nuxt has launched the Nuxt Agent (Beta) on nuxt.com, a first-party AI agent replacing the third-party Kapa AI widget. Built on top of the Nuxt MCP server, it provides grounded answers from official docs, modules catalog, blog, and deployment guides. The agent renders rich UI cards for modules, templates, and hosting providers, generates StackBlitz playground links, and streams responses end-to-end. Technically, it uses AI SDK v6 with claude-sonnet-4.6, Drizzle ORM for chat persistence, and a Nitro server handler. A key architectural decision is sharing the same MCP server with external AI assistants like Cursor and Claude Desktop. Rate limiting is set at 20 messages/day per IP. The full source code is open on GitHub.

Google Rapid Bucket delivers 23% PyTorch training speedup via gRPC-backed Colossus storage

Google is announcing Rapid Bucket, a high-performance object storage solution for Google Cloud that integrates with PyTorch via the gcsfs fsspec interface. By replacing REST-based storage access with persistent bidirectional gRPC streams backed by Google's Colossus architecture, it delivers a 23% training speed improvement and 4.8x read throughput gains. The integration requires zero code changes — developers only need to switch to a Rapid Bucket and upgrade gcsfs to version 2026.3.0. Benchmarks were run on 16 GKE nodes with A4 GPUs using PyTorch Lightning across 134M rows of data.

Uber launches hotel booking and credits AI tools for cutting development timelines in half

Uber announced hotel booking capabilities for US customers through a partnership with Expedia Group, giving access to over 700,000 hotels worldwide. Uber One members get 20% discounts on 10,000 hotels and 10% back in Uber Credits. The company's CTO credited agentic AI tools like Cursor for cutting development timelines in half, enabling features that previously took a year to ship to launch in months. Additional new features include travel mode with local guides, an Uber Eats room service hub, OpenTable restaurant reservations, and an 'Eats for the Way' snack pre-order for Uber Black rides.

IBM releases Granite 4.1 with 512K context and Apache 2.0 license

IBM releases Granite 4.1, a family of dense decoder-only LLMs (3B, 8B, 30B) trained on ~15 trillion tokens using a five-phase pre-training pipeline. Key highlights include long-context extension up to 512K tokens, SFT on ~4.1M LLM-as-Judge-curated samples, and a four-stage RL pipeline using on-policy GRPO with DAPO loss. Notably, the 8B instruct model matches or surpasses the previous Granite 4.0-H-Small (32B-A9B MoE) model. FP8 quantized variants are available for vLLM inference. All models are released under Apache 2.0.

Zed code editor reaches 1.0 with parallel agents, debugger, and Zed for Business

Study finds AI vision models give wildly inconsistent carb estimates across 27,000 queries

A preprint study tested four AI vision models (GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Pro, Gemini 3.1 Pro Preview) on carbohydrate estimation from food photos, submitting 13 images over 500 times each — 26,904 queries total. Key findings: all models showed significant within-model variability even at minimum temperature settings, with Gemini 2.5 Pro producing a 429g carb range (equivalent to 42.9 insulin units) for a single paella photo. Claude Sonnet 4.6 was the most consistent but still systematically wrong on some foods. All models overestimate carbs on average, biasing toward hypoglycemia risk. Critically, model-reported confidence scores had near-zero or negative correlation with actual accuracy, making them useless as safety filters. The study concludes that no current LLM is safe for unsupervised insulin dosing decisions.

Attackers abuse Kuse.ai platform to host phishing documents bypassing email filters

Threat actors abused Kuse.ai, a legitimate AI workplace platform, to host a phishing document as part of a Vendor Email Compromise (VEC) attack. Attackers uploaded a Markdown (.md) file containing a blurred fake document preview to Kuse's storage, then shared it via the platform's native link-sharing feature. The .md extension helped bypass common phishing filters. Victims were lured into clicking a link that redirected them to a fake Microsoft login page to harvest credentials. The attack combined a compromised vendor mailbox, platform legitimacy, and image manipulation for multi-layered social engineering. Recommendations include phishing-resistant MFA (FIDO2/WebAuthn), time-of-click URL inspection, VEC-aware security training, and restricting AI platform sharing features.

Railway ships soft-delete and scoped MCP tokens after AI agent wipes production database

An AI agent accessed a Railway API token stored on disk and called a volumeDelete mutation, wiping a customer's production database. Railway has since recovered the data and shipped fixes: all deletes now soft-delete for 48 hours (matching the dashboard behavior), backup cascading deletes are delayed, and the team is improving token scoping UX. The post outlines existing guardrails (staged changes, environment RBAC, project deletion grace periods) and new agent-friendly surfaces: a built-in Railway Agent with staged destructive operations, an improved CLI, and a Remote MCP Server at mcp.railway.com that uses short-lived tokens and scoped consent instead of long-lived disk tokens.

Canonical details how Ubuntu 26.04 LTS will natively package all three GPU compute stacks

Oracle NetSuite embeds ERP-specific skills into AI coding assistants via SuiteCloud Agent Skills

Oracle NetSuite is introducing SuiteCloud Agent Skills, which embed NetSuite-specific knowledge into AI coding assistants to help developers customize its ERP platform more efficiently. The skills cover UI framework references, permission codes, SuiteScript fields, OWASP security guidance, and migration tools from SuiteScript 1.0 to 2.1. NetSuite plans to roll out this guidance across more than 25 AI coding platforms. Analysts note the potential to reduce friction for enterprise developers, but caution that adoption may be slow due to sunk costs in existing development practices, token economics, and the risk of breaking tightly coupled ERP workflows. Sandboxed testing is recommended to catch code hallucinations and business logic issues.

Mistral AI launches Workflows for stateful enterprise AI orchestration in public preview

Mistral AI has launched Workflows, an orchestration layer for enterprise AI now in public preview. Built on top of Temporal and integrated into Mistral's Studio platform, it lets developers define multi-step AI processes in Python with stateful execution, fault tolerance, retry policies, and human-in-the-loop approval checkpoints. The architecture separates control and data planes, keeping execution workers and data within the customer's environment. It targets the gap between AI experimentation and reliable production deployment, addressing issues like pipeline failures, timeouts, and lack of audit trails in regulated environments. Early community reactions are cautiously optimistic, noting that lower-level reliability challenges around GPU utilization and real traffic handling remain unsolved.