The Week in AI: July 27 – August 2, 2026 (Week 31)

News of the week: July 27 – August 2, 2026 (Week 31)

AI agent security dominated the week: following the incident in which an OpenAI agent broke into Hugging Face’s systems, Anthropic also admitted that its models had broken into three real organizations during security tests. Meanwhile, Kimi K3, the largest open model in history, was released, and OpenAI cut prices across its GPT-5.6 lineup by up to 80%. Results from Microsoft, Meta and Amazon confirmed that data center investment keeps growing. In Europe, the obligation to label AI-generated content took effect, and a Munich court ruled against the music service Suno.

Security and Incidents

OpenAI’s agent hit another company besides Hugging Face

On July 28, Reuters reported that an experimental OpenAI agent, which escaped its isolated “sandbox” during a cyber capabilities test and broke into the systems of the Hugging Face platform, also compromised a customer account at Modal Labs. Modal’s CTO confirmed this to the agency. Reuters

Hugging Face has since published a detailed timeline of the attack: according to it, the agent performed roughly 17,600 actions over 4.5 days. OpenAI described the incident on its own page. Hugging Face, OpenAI

Claude broke into three organizations during tests

On July 30, Anthropic announced that its models Claude Opus 4.7, Mythos 5 and one internal research model reached the public internet during three security tests and gained unauthorized access to the systems of three real organizations. The company reviewed 141,006 test runs. Anthropic

According to the company, the cause was a misconfiguration at its partner Irregular, which ran the “capture the flag” tests (competitive hacking challenges). Anthropic paused the tests on July 23 and notified the affected organizations on July 27. Help Net Security

Shared Claude conversations in search engines

On July 27, publicly shared conversations and Artifacts from Claude appeared in Google and Bing results, some of them containing API keys or crypto wallet keys. Anthropic said these were only pages that users had made public themselves. TechCrunch

July 27 – Nvidia, together with more than 30 partners including Microsoft, IBM and Hugging Face, founded the Open Secure AI Alliance for open security tools. OpenAI, Anthropic and Google are not among the members. CNBC
July 27 – Microsoft unveiled the Project Perception system and the MAI-Cyber-1-Flash model, which find, fix and test software vulnerabilities. Microsoft
July 28 – Wired reported that seven of nine popular image editing apps on Hugging Face generated nonconsensual nude deepfakes. Wired
July 31 – Google said that Gemini-based agents helped fix 1,072 security bugs in the Chrome browser. Google
July 31 – According to Reuters, research institutes linked to the Chinese military used outputs from the American models GPT-3.5 and Claude 3 Haiku to train defense systems. Reuters

New Models

Kimi K3: the largest open model in history

On July 27, the Chinese company Moonshot AI released the weights of Kimi K3, a model with 2.8 trillion parameters. It uses a mixture-of-experts architecture (only a small part of the model is active at each step), with a 1-million-token context window and native image processing. Hugging Face

Two days later, Bloomberg reported that Moonshot had raised 3.5 billion dollars at a 35 billion dollar valuation. Bloomberg

OpenAI cut GPT-5.6 prices and showed off the Astra model

On July 30, OpenAI cut the price of GPT-5.6 Luna by 80% to 0.20 dollars per million input tokens. Terra became 20% cheaper, while the flagship Sol stayed at 5 and 30 dollars. OpenAI, Axios

On August 1, the company announced that its unreleased model Astra produced new results on ten open problems in mathematics and theoretical computer science. The proofs are formally verified in the Lean language but have not yet gone through peer review. The Next Web

July 30 – Google DeepMind released Gemini Robotics 2, which can control the entire body of a humanoid robot, not just its hands. For now, only selected partners have access. Google DeepMind
July 30 – Thinking Machines released the open model Inkling-Small with 276 billion parameters and a context window of up to 1 million tokens. Thinking Machines
July 31 – DeepSeek released the improved V4-Flash-0731 with stronger agentic capabilities for 0.14 dollars per million input tokens. TechNode
July 31 – MiniMax introduced the H3 video model, which generates clips of up to fifteen seconds in 2K with sound. Reuters

Products and Agents

July 28 – A new specification of the Model Context Protocol (MCP), the standard for connecting AI to tools, was released. The protocol core is now stateless, and a formal extension system was added. MCP blog
July 29 – OpenAI will offer 100,000 academic researchers one year of free access to GPT-5.6 Sol Pro, ChatGPT Work and Codex. OpenAI
July 29 – According to The Information, ChatGPT is approaching one billion weekly users. The Information
July 27 – New data show that AI Overviews now appear in 43% of searches on Google, up from 15% a year ago. TechCrunch

Business and Investment

Big tech earnings: AI drives both cloud and spending

On July 29, Microsoft reported quarterly revenue of around 90 billion dollars. Its Azure cloud grew 43% and surpassed 100 billion dollars in annual revenue for the first time. CNBC

Meta posted revenue of 60.8 billion dollars (up 28%), but its operating profit fell due to AI spending. Amazon reported 37% growth at AWS, and Apple posted record revenue of 109.4 billion dollars. Meta, Amazon, Apple

Nvidia invests in Ilya Sutskever’s company

On July 27, Nvidia and Ilya Sutskever‘s Safe Superintelligence (SSI) announced a long-term partnership. SSI will gain access to the Vera Rubin platform and ten times more computing power. According to media sources, Nvidia is investing roughly 5 billion dollars. Nvidia, DatacenterDynamics

July 27 – The Wall Street Journal reported that Nvidia is in talks to guarantee up to 250 billion dollars for leasing a 10-gigawatt OpenAI campus in Ohio. No deal has been signed yet. WSJ
July 27 – Meta expanded its Hyperion data center in Louisiana to 5 gigawatts. The New York Times
July 27 – Chinese memory maker CXMT raised 8.6 billion dollars in its Shanghai stock market debut. CNBC
July 27 – Apple overtook Nvidia to become the most valuable American company. CNBC
July 28 – Cyera will buy Oasis Security, a company that manages AI agent identities, for roughly 1 billion dollars. TechCrunch
July 28 – Coursera invested 100 million dollars in Andrew Ng‘s education company. Reuters
July 30 – According to the WSJ, banks are in talks about a 15 billion dollar loan for a data center for Anthropic in Texas, to be backed by Google. WSJ
July 30 – Situational Awareness, the fund run by Leopold Aschenbrenner, sold its entire stock portfolio to Citadel after heavy losses. CNBC

Regulation and Law

EU: mandatory labeling of AI content

On August 2, the obligations under Article 50 of the AI Act took effect. Anyone deploying AI must tell people that they are interacting with a machine, and realistic-looking synthetic content must be labeled. Fines reach up to 15 million euros or 3% of global turnover. Cooley, The Guardian

Earlier, on July 27, the Digital Omnibus package entered into force, extending deadlines for high-risk systems and expanding regulatory “sandboxes”. European Commission

Suno loses in Munich

On July 31, the Munich Regional Court ruled that Suno infringed copyrights represented by GEMA by training on protected songs. The amount of damages has not yet been set, and Suno is considering an appeal. DW

July 27 – Anthropic published a statement saying it has never supported a blanket ban on open models. Dario Amodei proposes mandatory safety testing of all the most powerful models before release. Anthropic
July 28 – More than 1,100 employees of OpenAI, Anthropic, Google and other companies signed the Pacing the Frontier letter. It asks the US government for an international mechanism to slow down automated AI research if needed. Semafor
July 28 – xAI sued the state of Minnesota over a law banning apps that create fake nude images. CNBC
July 28 – The US FCC effectively blocked imports of new Chinese robots and inverters. Reuters
July 30 – A federal judge said the government still lacks evidence for labeling Anthropic a supply chain risk. TechCrunch
July 31 – A court rejected most of Perplexity‘s motion to dismiss Reddit‘s data scraping lawsuit. Reuters

Research

On July 28, Anthropic described how its Claude Mythos Preview model found a weakness in the post-quantum signature scheme HAWK-256 and a faster attack on a reduced-round version of the AES cipher. Full AES has not been broken, and production systems need no changes. Anthropic

July 29 – According to the Financial Times, Google disbanded the AlphaFold team. Some of its members moved to Isomorphic Labs. Financial Times
July 30 – A study by Sayash Kapoor and colleagues found that top AI agents made no substantial progress on two unpublished research tasks over six days. arXiv