Week in AI: May 11-17, 2026 (Week 20)

This week’s news: May 11-17, 2026 (Week 20)

The week was dominated by cybersecurity and big money. OpenAI launched its Daybreak defense platform and set up a $4 billion company to roll out AI directly inside large enterprises. According to Bloomberg, Anthropic is in talks over a $30 billion investment at a valuation above $900 billion, and it struck partnerships with the Gates Foundation and PwC. Chipmaker Cerebras went public in the largest US IPO of the year. In Oakland, the evidence phase of Elon Musk’s lawsuit against OpenAI came to a close, and Google for the first time caught attackers who had used artificial intelligence to find an exploitable bug.

Products and Agents

OpenAI Daybreak: An Answer to Claude Mythos

On Monday, May 11, OpenAI unveiled its Daybreak initiative. It combines the GPT-5.5 family of models, the Codex Security tool and a network of partners, and is designed to help defenders find vulnerabilities in software, verify fixes and ship patches faster.

A vulnerability is a flaw in a program that an attacker can exploit. Daybreak can scan repositories (source code storage), propose fixes and test them. Commentators see it as a direct response to the Claude Mythos model and the Glasswing project from Anthropic (Engadget, TechRadar).

Codex on Mobile and ChatGPT for Personal Finance

On Thursday, May 14, OpenAI added its coding agent Codex to the ChatGPT mobile app for iOS and Android. There is no coding on the phone itself: users connect to a session running on a Mac, follow the agent’s work, approve commands and assign new tasks.

During the trial period, the feature is available on all plans, including the free one. According to the company, Codex is used by more than 4 million people every week (9to5Mac).

On Friday, May 15, it followed up with a preview of personal finance tools for ChatGPT Pro subscribers in the US. Through the Plaid service, users can connect accounts at more than 12,000 banks and brokerages and get a breakdown of their spending or help with planning (TechCrunch).

Anthropic Targets Small Businesses and AWS

On Wednesday, May 13, Anthropic launched its Claude for Small Business package. It includes integrations with tools such as QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace and Microsoft 365, plus about fifteen ready-made workflows for accounting, HR, marketing and operations.

It is turned on with a single toggle in the Claude Cowork app. Claude can then, for example, match PayPal payments with receivables in the books and prepare payment reminders for approval (Axios).

Back on May 11, Amazon Web Services announced the general availability of Claude Platform on AWS. Companies get Anthropic’s native platform directly through their AWS account, using its sign-in and billing. The service is operated by Anthropic, and data is processed outside the AWS security boundary (AWS).

Other News

xAI released an early beta of Grok Build, a command-line coding agent, on May 14. For now it is available only to SuperGrok Heavy subscribers and competes with Claude Code and Codex (Engadget).
Mark Zuckerberg introduced Incognito Chat for Meta AI in the Meta AI app and in WhatsApp on May 13. According to Meta, conversations are not stored on its servers, and processing runs in an isolated trusted execution environment that the company cannot see into. An independent audit is still missing (MacRumors).
Google ahead of its I/O conference: code in the beta version of the Google app shows that the agent in Gemini is being renamed from “Gemini Agent” to “Gemini Spark” (May 14, 9to5Google).

Business and Investment

OpenAI Sets Up an AI Deployment Company and Buys Tomoro

On May 11, OpenAI launched a new entity, the OpenAI Deployment Company, with initial capital of more than $4 billion. The round was led by TPG, Advent, Bain Capital and Brookfield, while OpenAI retains a majority stake and control.

The company will send engineers directly into enterprises to deploy the latest models there. Right from the start, it is acquiring the Scottish consultancy Tomoro, which brings about 150 specialists; the price was not disclosed (The Next Web, Bloomberg Law).

OpenAI and Microsoft Rewrite Their Relationship

According to The Information, whose report was picked up by Reuters on May 12, OpenAI and Microsoft have agreed to cap revenue sharing at $38 billion. OpenAI will keep paying Microsoft the same percentage until 2030, but the total amount will be capped (Reuters via Yahoo Finance).

A day later, Reuters reported that Microsoft is looking for acquisitions for life after OpenAI. Among others, it is in talks with the startup Inception, which develops diffusion language models: instead of generating text word by word, they refine the entire output at once (Reuters via The Star).

Anthropic: Record Valuation, the Gates Foundation and PwC

According to Bloomberg on May 12, Anthropic is in talks to raise at least $30 billion at a valuation above $900 billion. That would put it ahead of OpenAI for the first time (Bloomberg).

On May 13, the company also warned that transfers of its shares without board approval are void. It named eight unauthorized intermediaries, including the Hiive and Forge platforms (Axios).

On Thursday, May 14, Anthropic and the Bill & Melinda Gates Foundation announced a four-year, $200 million partnership. Funding, Claude credits and engineering support will go toward global health, life sciences, education and economic mobility. The resulting tools are to be freely available (The Next Web).

The same day, consulting firm PwC expanded its partnership with Anthropic. Claude Code and Claude Cowork will go first to US teams and then to other employees, and the firm will train and certify 30,000 people (PwC, SiliconANGLE).

Cerebras in the Biggest IPO of the Year

AI chipmaker Cerebras listed on the Nasdaq on May 14. At $185 per share, it raised about $5.55 billion, the largest US IPO so far this year.

The shares opened at $350 and closed their first day at $311.07, roughly 68% higher. According to reports, demand exceeded supply more than twentyfold (Bloomberg Law, Yahoo Finance).

Samsung workers threatened an 18-day strike over how profits from the AI-driven memory chip boom are shared. Employees in the logic chip and foundry divisions feel left out (May 15, Reuters).

Regulation, Politics and Courts

Musk v. OpenAI Heads to the Jury

On Thursday, May 14, closing arguments were delivered in federal court in Oakland in Elon Musk‘s lawsuit against OpenAI. Musk claims the company betrayed its original nonprofit mission. OpenAI counters that the lawsuit proved none of its allegations.

The nine-member jury is set to begin deliberating the following Monday, and its verdict will be only advisory for Judge Yvonne Gonzalez Rogers. Witnesses during the week included Sam Altman and Reid Hoffman; Musk missed the closing arguments because he had flown to China with President Trump (Local News Matters, The Star).

Trump and Xi Jinping on AI Rules

After the summit in Beijing, Donald Trump said on May 15 that he had discussed possible cooperation on “guardrails” for artificial intelligence with Xi Jinping. No specific signed AI agreement came out of the talks.

They also discussed Nvidia H200 chips, whose sale to China the US had already approved in December. US Trade Representative Jamieson Greer said that purchases are now up to the Chinese side (Bloomberg).

Security and Incidents

Google Catches the First Exploitable Bug Found with AI

The Google Threat Intelligence Group announced on May 11 that it had disrupted a criminal group’s plan for a mass attack. According to the team, the attackers used artificial intelligence to find and weaponize a previously unknown flaw, a so-called zero-day, that made it possible to bypass two-factor authentication in a popular open source server administration tool.

Traces in the code pointed to the use of AI, such as a made-up severity rating. Google worked with the vendor to fix the bug before the attack could get underway (Fortune).

OpenAI Hit by the TanStack Package Attack

On May 13 and 14, OpenAI confirmed that a supply chain attack using tampered versions of TanStack libraries in the npm package manager had compromised two employee devices. According to TechCrunch, the attackers made off with a limited amount of internal credentials.

The company says user data, production systems and software were not affected. As a precaution, however, it rotated its code-signing certificates, and Mac users need to update the ChatGPT, Codex and Atlas apps (TechCrunch, The Register).

Research

AI Cyber Capabilities Are Growing Faster Than Expected

The UK AI Security Institute published new measurements on May 14. By its earlier estimate, the length of cyber tasks that models can complete on their own was doubling roughly every 4.7 months; as recently as November, that figure was eight months.

According to the institute, the Claude Mythos Preview and GPT-5.5 models both significantly outpaced these trends. It is not yet clear whether this marks the start of an even faster trend (Help Net Security, The Register).