AI News May 11–15, 2026: Musk v. OpenAI, Anthropic Mythos, Grok Build CLI

May 15, 2026 Regulation

Musk v. OpenAI heads to closing arguments — Sutskever, Nadella and Altman among trial witnesses

Closing arguments in Elon Musk's lawsuit against OpenAI took place on 14–15 May, with Musk's counsel using a "wooden bridge" analogy to challenge Sam Altman's truthfulness. The trial heard testimony from co-founder Ilya Sutskever — whose statements were read as cutting both ways on the nonprofit-mission question — and from Microsoft CEO Satya Nadella on the structure of the partnership.

Why it matters: A judgment in either direction would set a benchmark for governance disputes at frontier AI labs and the enforceability of nonprofit-mission promises against later for-profit pivots.

Google says it disrupted the first known AI-assisted zero-day exploit — used to bypass 2FA

Google's Threat Intelligence Group (GTIG) reported what it describes as the first identified case of a prominent cybercrime group using a large language model to discover and weaponise a zero-day vulnerability. The target was an open-source web-based system administration tool, and the exploit chain enabled two-factor authentication bypass. Google states it disrupted the campaign before mass exploitation, and noted hallucinated artefacts in the code (including invented CVSS scores) consistent with AI assistance.

Why it matters: This is the clearest public confirmation yet that criminal groups are operationalising AI for vulnerability research, accelerating the shift in the offence–defence balance that has been forecast for two years.

Anthropic Mythos credited with finding two macOS bugs — researchers drove to Apple HQ to disclose

Security researchers reported that Anthropic's still-unreleased Mythos vulnerability-research model helped surface two exploitable bugs in macOS, including one affecting M5-era protections. The Wall Street Journal describes Mythos as a model shared under restricted access with selected partners through what Anthropic has previously discussed as Project Glasswing. Researchers disclosed the findings to Apple in person.

Why it matters: A vendor-supplied AI finding live OS-level bugs in days rather than months is a tangible demonstration of why frontier cyber-models are now treated as a controlled-distribution category by labs and governments.

xAI launches Grok Build CLI — a terminal-based coding agent for SuperGrok Heavy subscribers

xAI opened an early beta of Grok Build, a terminal-based coding agent aimed at professional software engineering. The product supports agentic workflows with subagents, planning and tool use, and is gated to SuperGrok Heavy subscribers — a tier widely reported at around $300 per month. xAI is positioning Grok Build as a competitor to Claude Code and OpenAI Codex.

Why it matters: The terminal-agent category — once a Claude-and-OpenAI race — now has three serious entrants, signalling that the developer console has become the primary battleground for AI agent UX.

Thinking Machines releases Interaction Models — voice AI that can wait, interrupt and watch posture

Thinking Machines published a research demo of its Interaction Models, a system focused on natural real-time voice conversation. The model can pause and wait, interrupt politely, count items in view, observe posture from camera input, and redirect requests it judges unsafe. The release is the lab's first major public capability demonstration.

Why it matters: Real-time multimodal interaction has been a stated goal of every major lab since GPT-4o; Thinking Machines arriving with a competitive demo broadens the frontier beyond the incumbents.

OpenAI launches Daybreak — a request-driven cybersecurity scanner for customer systems

OpenAI announced Daybreak, a cybersecurity offering that scans customer systems for vulnerabilities on request. OpenAI is positioning the product as a more constrained, opt-in counterpart to broader cyber-capable models — in implicit contrast to Anthropic's Mythos / Glasswing approach — with use gated to verified defenders.

Why it matters: Frontier labs are now openly differentiating themselves on cyber-distribution policy, not just capability, as the dual-use risk of vulnerability-finding AI becomes a regulatory pressure point.

DeepMind unveils AI Pointer — a context-aware cursor for editing documents and images by voice

Google DeepMind introduced AI Pointer, a system that lets a user highlight content on screen and issue voice instructions to edit it — applied to documents, images, lists and other interface elements. The pointer interprets selection plus speech as a unified intent, removing the need to switch between dialog chat and direct manipulation.

Why it matters: Embedding AI control at the cursor level rather than in a sidebar reframes how operating systems and applications expose AI affordances to users — a fundamentally different UX from chat-window assistants.

Krea releases Krea 2 image model — finer style control and weighted reference mood boards

Krea released Krea 2, its next-generation image model. The update emphasises style control, weighted reference images and mood-board-driven generation, giving users explicit knobs over how much each input image influences the output.

Why it matters: Style-controllable generation with explicit reference weighting is the feature professional creatives have been asking for, and a useful counter-positioning for smaller image labs against the bigger generalist models.

World Labs ships image-to-3D — single photo becomes an interactive scene with physics and audio

World Labs released an open-source tool that converts a single image into a navigable 3D environment with movable objects, physics, lighting and audio. The release lowers the barrier for hobbyists and indie developers to produce playable 3D scenes from reference photography.

Why it matters: Single-image-to-scene with physics and audio is a key step toward AI-generated game and simulation content, and the open-source release accelerates downstream tooling.

May 13, 2026 Product

Anthropic ships Agent View in Claude Code — one interface for many concurrent agents

Anthropic introduced Agent View, a consolidated interface inside Claude Code for managing multiple concurrent coding agents from a single view rather than juggling terminal windows. The release follows April's Claude Code desktop redesign that first pushed Anthropic's parallel-agent workflow concept.

Why it matters: Agent UX is becoming the differentiator. As parallel agents become routine, the bottleneck shifts from individual capability to how well the user can supervise a fleet of them.

Anthropic overtakes OpenAI in business adoption — 34.4% versus 32.3%

Ramp's leading-indicators report shows Anthropic surpassing OpenAI on business adoption among Ramp customers for the first time, with Anthropic at 34.4% and OpenAI at 32.3%. The shift reflects Claude's traction with enterprise developer and legal workloads through 2026.

Why it matters: Crossing OpenAI in business adoption is a symbolic milestone for Anthropic and a useful data point for anyone tracking how AI procurement is splitting between labs at the enterprise tier.

Anthropic launches Claude for Legal and packaged agents for small business

Anthropic continued its industry-specific push with Claude for the Legal Industry — tailored workflows, citation handling and document review aimed at law firms — and a separate Claude for Small Business release that ships ready-made agents for common operations tasks.

Why it matters: Verticalised AI products mark a clear shift away from generic chat toward sector-specific configurations — and intensify competition with Harvey, Hebbia and other domain-focused players.

OpenAI brings Codex to mobile — remote control of a desktop coding session from a phone

OpenAI expanded Codex with mobile access, allowing users to control a paired computer, browse files and continue coding tasks remotely from a phone. The mobile client is positioned for hand-off between desktop work sessions rather than full coding on a small screen.

Why it matters: Mobile control of long-running desktop agents normalises the workflow of starting an AI task and checking on it from anywhere — a usage pattern that increasingly looks like remote project management rather than coding.

Google launches Gemini Intelligence on Android and unveils "Googlebook"

Google announced a major Android update with deeper Gemini integration ("Gemini Intelligence"), and introduced Googlebook — positioned as the AI-first successor to Chromebook. Both moves are timed to align with Google I/O on 19–20 May.

Why it matters: Google is rebuilding its consumer hardware story around AI rather than the browser, in direct response to Apple Intelligence and the rise of AI-native laptops.

Notion launches a developer platform — CLI, workers, agents SDK and webhooks

Notion announced a developer platform that includes a CLI, background workers, an agents SDK and webhooks. The release moves Notion from an integrations-friendly product surface toward a programmable runtime that third-party developers can extend with native primitives.

Why it matters: Notion repositioning as an agent platform reflects a broader pattern: productivity tools competing on how easily their data and surfaces can be wrapped in AI workflows.

Meta adds Incognito chat to WhatsApp Meta AI and broadens Muse Spark rollout

Meta rolled out an Incognito mode for Meta AI conversations inside WhatsApp, where messages are not retained or used for personalisation. Separately, Meta expanded access to Muse Spark, the company's most capable Superintelligence Labs model.

Why it matters: Privacy-mode AI inside the world's biggest messenger is a meaningful concession to user concerns about chat-as-training-data — and Muse Spark expansion widens Meta's user-facing AI footprint.

Claude Code raises usage caps 50% — but new third-party credit system draws pushback

Anthropic increased Claude Code usage limits by 50% on a temporary basis and introduced a new credit system for third-party tool calls. Heavy users on developer forums and X criticised the credit model as opaque pricing on top of a subscription, with some threatening to cancel.

Why it matters: Capacity pressure and pricing changes for coding agents are now a recurring story — they shape who actually keeps using these tools for production work versus who churns to a competitor each quarter.

Claude reportedly helps recover 11-year-old Bitcoin wallet — 5 BTC restored from old hard drive

A user posted that Claude helped recover roughly 5 BTC from files on an 11-year-old hard drive, with the recovered wallet now worth several hundred thousand dollars at current prices. The thread describes feeding fragments and clues from old drives into Claude to reconstruct keys and metadata.

Why it matters: Beyond the headline, the underlying capability — making sense of partial, decade-old technical artefacts — is a non-obvious use case for large models that quietly extends what an individual can recover on their own.

Unitree unveils GD01 — a manned transformable "mecha" robot at roughly $650,000

Chinese robotics firm Unitree revealed the GD01, marketed as the world's first production-ready manned transformable robotic vehicle. The unit transforms between bipedal and quadrupedal modes, carries a human rider with a combined mass around 500 kg, and is targeted at civilian and obstacle-traversal use cases. Starting price is reported around 3.9 million yuan (~US$650,000).

Why it matters: Manned reconfigurable robotics moves China's robot industry further into the consumer-spectacle category — but the price point and homologation questions mean adoption will be limited to demonstrations and well-funded enthusiasts for now.

May 11, 2026 Incident

Vercel discloses security incident — third-party AI tool account compromise behind unauthorised access

Vercel published an incident report describing unauthorised access traced to the compromise of a third-party AI tool account and a Vercel employee account. The company said a limited subset of non-sensitive customer environment variables was affected and that core production systems were not breached.

Why it matters: The supply chain for developer-facing AI tooling is becoming a meaningful attack surface — every connected IDE plugin or agent is now a potential entry point into the customer's cloud platform.

Figure runs a multi-day humanoid livestream — robots sort packages for 34+ hours continuously

Figure ran a public livestream of its humanoid robots performing warehouse-style package sorting continuously for over 34 hours. The demonstration is timed alongside Figure's reported Series C funding at a $39B post-money valuation and is positioned as a showcase of real-world endurance rather than choreographed demos.

Why it matters: Multi-day continuous operation is the metric that matters for real warehouse adoption — and a public livestream is the kind of evidence that buyers (and competitors) cannot easily dismiss.

Rivian rolls out in-car AI assistant — knows the vehicle, reads texts, controls features by voice

Rivian shipped a software update introducing an in-car AI assistant that understands the vehicle's systems, can answer questions from the owner's manual, read incoming text messages aloud, and control vehicle features via voice. The assistant is part of Rivian's broader push to make over-the-air software a core differentiator.

Why it matters: EV makers are racing to embed AI assistants that go beyond entertainment into the vehicle's mechanical and safety knowledge — a category Tesla, Lucid and the Chinese manufacturers are all chasing.

Digg relaunches as an AI-curated trending feed at di.gg/ai

Digg returned in an AI-curated form, surfacing trending AI news ranked by signals from prominent voices on X. The new product is pitched at people who used to scan Hacker News or Techmeme but want a more opinionated, personality-driven feed.

Why it matters: Algorithmic news aggregation is being rebuilt on top of LLMs and social-graph signals — and a once-dominant 2000s brand returning as an AI product is a useful reminder of how the news-discovery layer keeps getting rewritten.

Real Monet painting mistaken for "AI slop" — viral Instagram post highlights AI-detection fatigue

A widely shared Instagram post showed a genuine Monet painting being criticised as "AI-generated slop" and dismissed as soulless by commenters. The thread underlined how quickly audiences now reach for the AI-detection reflex when assessing visual art online.

Why it matters: Defensive AI-pattern matching by audiences is starting to misfire against legitimate human work — a cultural consequence of generative image proliferation that has practical implications for artists and platforms.

← Previous Week (Apr 14–18) All Weeks Next Week (May 18–22) →

AI News This Week