The AI Upload Weekly Digest

Cognition, the startup behind autonomous AI software engineer Devin, has raised over $1 billion at a $25 billion pre-money valuation, a dramatic jump from its $10.2 billion post-money valuation just eight months ago. The round was led by Lux Capital and General Catalyst, with participation from Founders Fund, 8VC, Ribbit Capital, Atreides, and Layer Global.

The raise signals continued investor confidence in independent AI coding startups, even as major model makers including Anthropic, OpenAI, and Google have moved aggressively into the space. Cognition reports $492 million in annualized revenue run-rate, with enterprise usage of Devin growing 50% month-over-month for the past six months. Customers include Mercedes-Benz, NASA, Goldman Sachs, and Santander. For legal professionals tracking AI's expanding role in enterprise software development, the scale and speed of this growth are worth watching.

In a remarkable two-hour span on Wednesday, the AI industry generated a cascade of major developments touching nearly every dimension of the sector's growth. OpenAI announced that a general-purpose reasoning model independently solved a geometry problem that had eluded mathematicians for 80 years. Anthropic reported revenue on track to exceed $10.9 billion in Q2, putting the company ahead of its own profitability projections by two years, while separately committing roughly $1.25 billion per month through 2029 for access to SpaceX's Colossus supercomputing infrastructure. Nvidia posted $81.6 billion in quarterly revenue, with data center operations alone accounting for $75.2 billion.

On the policy front, President Trump summoned leading tech CEOs to Washington for the signing of an AI executive order expected to establish a voluntary framework requiring labs to share new models with the government 90 days before public release. The same day, Meta announced layoffs of approximately 8,000 workers even as it accelerates AI investment. Recent polling cited in the article found that 70% of Americans believe AI is advancing too quickly, and nearly two-thirds doubt the technology will produce broadly shared economic benefits.

Google CEO Sundar Pichai sat down with The Verge's Nilay Patel following Google I/O to discuss the company's AI-driven restructuring, the future of Search, and what those changes mean for the broader web. Pichai described reorganizing Google around a unified AI infrastructure, merging Brain and DeepMind into Google DeepMind, and centralizing AI infrastructure under a dedicated SVP, moves he says were prompted by the competitive pressure following ChatGPT's rise.

Of particular note for publishers and content creators: Pichai acknowledged the growing reality of what Patel has termed "Google Zero," the prospect of search traffic to third-party websites falling to nothing as Google answers queries directly. Major publishers, including Condé Nast, are reportedly already planning for a zero-traffic future. Google is also training models on YouTube content and restructuring YouTube search to surface video segments directly, a shift likely to generate friction with creators similar to ongoing tensions with publishers.

A federal judge in California has allowed a major copyright lawsuit against Chinese AI company MiniMax to move forward. US District Judge Stanley Blumenfeld rejected MiniMax's motion to dismiss, finding that the court has jurisdiction over the company and that the studios' complaint states plausible claims for copyright infringement.

Walt Disney, Comcast's Universal, and Warner Bros Discovery sued MiniMax last year, alleging the company trained its Hailuo image-and-video-generating AI system on their copyrighted content without permission. The studios further alleged that MiniMax used characters from the Marvel universe, Star Wars, and other franchises to market Hailuo as a "Hollywood studio in your pocket." MiniMax had argued the California court lacked jurisdiction because it conducts its business in China, but Judge Blumenfeld found evidence that the Hailuo app was being offered to US users.

The ruling is a significant procedural development in the broader wave of AI copyright litigation. The same studios have filed parallel suits against AI company Midjourney, and the case joins a growing body of litigation from media companies, authors, and news organizations challenging the use of their content to train AI systems.

Harvey, the legal AI company valued at $11 billion, released the Legal Agent Benchmark (LAB) on May 6, an open-source evaluation framework designed to measure how well AI agents perform extended legal work. Unlike existing benchmarks such as LegalBench and CUAD, which test discrete reasoning tasks, LAB focuses on longer-horizon assignments modeled on partner-to-associate delegations. The initial release includes more than 1,200 tasks across 24 practice areas, graded against over 75,000 expert-written rubric criteria. Notably, Harvey launched without a leaderboard, stating it will collaborate with research partners to establish baseline results and submission standards before publishing rankings.

LAB uses "all-pass" grading, meaning a task is complete only if every rubric criterion is met, reflecting the reality that missing a single material issue in a deal memo can have serious consequences. Harvey acknowledges contributions from major AI labs including Anthropic, OpenAI, and Google DeepMind. For law firms, the benchmark could provide a shared framework for vendor evaluation beyond demo-driven procurement decisions.

The article raises a significant caveat: LAB was built by a commercial market participant, meaning Harvey's team shaped the definitions of acceptable legal work product. Questions also remain about whether the project will attract genuine outside contributors or function primarily as a vendor showcase.

A three-week federal trial in Oakland pitting Elon Musk against OpenAI and CEO Sam Altman ended Monday without a verdict on the merits. The jury dismissed Musk's lawsuit on procedural grounds, finding it missed a statutory deadline. Musk had accused Altman and co-founder Greg Brockman of betraying OpenAI's original nonprofit mission and unjustly enriching themselves; OpenAI countered that Musk had supported the for-profit conversion and filed suit only to disadvantage a competitor as he built his own AI company, xAI.

Though no verdict was reached, trial testimony revealed the financial pressures that drove OpenAI's transformation from a nonprofit founded in 2015 to a for-profit enterprise now valued at $852 billion. Internal emails and witness accounts, including testimony from Microsoft CTO Kevin Scott and former chief scientist Ilya Sutskever, documented how the capital demands of large-scale AI development left the organization with few alternatives to commercial investment. For legal professionals tracking AI governance and fiduciary obligations in emerging-technology organizations, the trial record offers a rare window into how structural decisions at a major AI company were actually made.

Google's annual I/O developer conference delivered a sweeping set of AI announcements centered on new models, agentic capabilities, and consumer-facing tools. The company launched Gemini 3.5 Flash, a model positioned to match flagship-level performance at faster speeds and lower cost, with general availability through the Gemini API and Google AI Studio. A new multimodal model, Gemini Omni, adds video generation and editing capabilities with built-in SynthID watermarking for content verification. Google also announced Gemini Spark, described as a 24/7 personal AI agent that can act autonomously on a user's behalf, currently in limited testing.

On the search and commerce side, Google reported AI Mode has surpassed one billion monthly users and announced a redesigned search interface supporting text, images, files, and video inputs. A new Universal Cart feature integrates shopping across Search, Gemini, YouTube, and Gmail, with background price-monitoring and checkout through Google Pay. For legal professionals and researchers, the expansion of Personal Intelligence in AI Mode to nearly 200 countries, with the ability to connect Gmail and Google Photos, raises notable questions about data integration and user control that the company says is built on transparency and user choice.

In late 2025, two tools reshaped how technically proficient developers interact with software: Anthropic's Opus 4.5 update to Claude Code and an open-source personal agent called OpenClaw, built by entrepreneur Peter Steinberger. Opus 4.5 could run for hours, manage teams of AI subagents, and, by Anthropic's own account, outscored every human candidate on the company's engineering hiring exam. Steinberger's OpenClaw, which gives an AI agent access to a user's apps, data, and even payment credentials to autonomously complete tasks, accumulated over 366,000 GitHub stars within months of release.

The article profiles the developers and early adopters driving this shift, including Anthropic engineer Boris Cherny, who built Claude Code after returning from rural Japan, and Y Combinator CEO Garry Tan, who estimated Claude Code multiplied his coding output to the equivalent of over 400 engineers. For legal professionals, the implications are significant: the same autonomous agent capabilities now transforming software development are architecturally similar to tools that could soon automate research, drafting, and document review workflows at comparable scale.

A critical vulnerability discovered in Starlette, an open source framework downloaded 325 million times per week, puts millions of AI agents and their underlying servers at serious risk. The flaw is described as trivial to exploit and affects not only Starlette itself but thousands of dependent projects, including FastAPI and other widely used Python frameworks. Security researchers warn that patches or mitigations are urgently needed given the scale of exposure.

The vulnerability is particularly consequential because Starlette underpins ASGI implementations that connect to MCP (model context protocol) servers, the infrastructure enabling AI agents from major providers to interact with external resources such as databases, email, and calendar accounts. Those MCP servers store credentials for each connected system, making a successful breach potentially devastating in scope. Legal professionals and law firms relying on AI tools integrated with sensitive client data and third-party services should monitor this situation closely.

Anthropic has announced plans to eventually release Mythos-class models to the general public, contingent on developing stronger safety guardrails. Mythos, the company's specialized security vulnerability detection AI, has been available only to vetted partners through "Project Glasswing" since its April announcement due to concerns that open access would enable cybercriminals to exploit software flaws at scale. Anthropic acknowledged in its project update that "at present, no company, including Anthropic, has developed safeguards strong enough to prevent such models from being misused and potentially causing severe harm," while offering no timeline for what it called a "near future" general release.

The scale of Mythos's findings underscores both the tool's power and the challenges it creates. Scanning more than 1,000 open-source projects, Mythos identified over 23,000 vulnerabilities total, including 6,202 rated high or critical severity. Among the confirmed critical flaws was a certificate-forging vulnerability in the wolfSSL cryptography library, used by billions of devices worldwide. Open-source maintainers have reportedly asked Anthropic to slow its disclosure rate due to capacity constraints, a dynamic that raises significant questions for legal and compliance professionals tracking software liability and coordinated vulnerability disclosure practices.

Anthropic has disclosed that Project Glasswing, its defensive cybersecurity initiative launched last month, has uncovered more than 10,000 high- or critical-severity vulnerabilities in widely used software. The effort gives approximately 50 vetted partners early access to Claude Mythos Preview, a frontier AI model designed to autonomously identify security flaws before malicious actors can exploit them. Of the vulnerabilities identified, 1,094 have been confirmed as high- or critical-severity true positives, 97 have been patched upstream, and 88 advisories have been issued. One notable find is a critical flaw in WolfSSL with a CVSS score of 9.1 that could allow certificate forgery.

Anthropic is urging software developers and network defenders to shorten patch cycles in response, warning that models with similar capabilities to Mythos Preview may become broadly available soon. The company has also launched a Cyber Verification Program allowing security professionals to use its models without standard guardrails for legitimate purposes such as penetration testing and red teaming. Mythos Preview and a comparable OpenAI model, GPT-5.5-Cyber, have not been released publicly due to concerns about the absence of adequate safeguards against large-scale misuse.

Microsoft has issued a warning about an active cryptojacking campaign that exploits AI chatbot interactions to steer users toward malicious software download sites. When users ask large language model-based tools for software recommendations, the chatbots surface links to attacker-controlled domains, an extension of traditional search engine optimization poisoning into AI-generated responses. More than 150 malicious domains have been identified in the campaign.

The operation deliberately targets systems with high-performance GPUs by impersonating legitimate utilities such as CrystalDiskInfo and HWMonitor, maximizing cryptocurrency mining yield per compromised device. Downloaded files install ScreenConnect remote access software alongside mining programs, while the malware configures persistence mechanisms and disables antivirus detection. Beyond financial gain, the threat actors use their remote access for potential data theft, lateral movement, and ransomware deployment.

Microsoft said it detected and blocked activity associated with the campaign. For legal professionals and law students relying on AI tools for research and software guidance, this campaign is a reminder that AI-generated recommendations are not inherently vetted for safety.

Hours before President Trump was set to sign an executive order requiring a government safety review of new AI models before release, he abruptly reversed course. Trump cited competition with China as his rationale, stating he did not want to do anything to slow the United States' lead in AI development. The reversal followed private lobbying by tech billionaires including Elon Musk, Mark Zuckerberg, and former White House AI adviser David Sacks, who reportedly warned the president the order would harm the economy and benefit China.

The episode was triggered by Anthropic's announcement of its Claude Mythos model, which the company withheld from public release due to its ability to identify vulnerabilities in computer code. The model prompted concern from governments worldwide and briefly appeared to shift the administration toward considering some oversight. The proposed order, however, was already notably limited, explicitly prohibiting any mandatory licensing or preclearance requirements for AI model releases.

The outcome signals that meaningful federal AI regulation remains a distant prospect. A draft of the shelved order confirmed it carried no legal enforcement power. Meanwhile, industry-aligned super PACs are amassing hundreds of millions of dollars to support anti-regulation candidates ahead of midterm elections, suggesting the tech industry's influence over AI policy is likely to strengthen rather than diminish.

Alex Bores, a New York state assemblyman who authored the RAISE Act placing restrictions on frontier AI model releases, entered the race for New York's 12th congressional district last October as a relative unknown. That changed when Leading the Future, a super PAC backed by OpenAI's Greg Brockman, Marc Andreessen, Joe Lonsdale, and Palantir executives, began running attack ads against him in December 2025, ultimately spending an estimated $2.4 million. Rather than sidelining his candidacy, the campaign elevated it.

The Bores campaign placed its first ad buy only on May 11th, nearly seven months into the race. Yet a recent Emerson College poll shows him trailing frontrunner Micah Lasher by just two points. Internal campaign polling indicates that voters who received negative information about Bores were actually more inclined to support him. Anthropic's $20 million donation to a pro-regulation PAC backing Bores then drew coverage from The New Yorker, The New York Times, and Politico, framing the race as a proxy war between the two AI labs.

For legal professionals tracking AI regulation and the role of money in tech policy, this race offers a striking case study in how industry opposition to regulatory candidates can backfire, and how AI governance is increasingly shaping electoral politics at the congressional level.

China is tightening its grip on top AI talent, with researchers, startup founders, and executives at private firms now reportedly subject to travel restrictions requiring government approval before going abroad. The restrictions have intensified following scrutiny of Meta's $2 billion acquisition of AI startup Manus, whose two co-founders have been barred from leaving the country while Beijing investigates whether the deal violates foreign investment rules. The co-founders are reportedly exploring options to unwind the acquisition, including raising approximately $1 billion to buy the company back from Meta.

The moves are part of a broader set of controls Beijing is deploying to protect its AI sector, including requiring government approval before firms like Moonshot AI, StepFun, and ByteDance can accept American capital. These developments come as the performance gap between leading U.S. and Chinese AI models has narrowed sharply, from roughly 31% in 2023 to just 2.7% as of March 2026, according to Stanford's latest index. For legal professionals tracking cross-border technology transactions, foreign investment review, and export controls, this story signals a rapidly shifting regulatory landscape with significant implications for international AI deals.

Pope Leo XIV has issued his first papal encyclical, "Magnifica Humanitas," a 42,300-word document addressing artificial intelligence and its effects on humanity. Drawing explicit parallels to his predecessor Leo XIII's 1891 labor encyclical "Rerum Novarum," the current pope frames AI as a potential threat to worker dignity and calls for government regulation of AI companies, algorithmic transparency, job protection measures, and a prohibition on lethal autonomous weapons systems. The document argues that moral judgment over life-and-death decisions cannot be delegated to AI systems.

The encyclical has drawn criticism from multiple directions. AI ethics researcher Timnit Gebru objected to the Vatican's decision to seat Anthropic co-founder Chris Olah alongside the pope at the document's presentation, arguing the church was platforming a company it should be scrutinizing. From the right, libertarian policy analyst Dean Ball contended the encyclical dodges the deeper philosophical challenge AI poses to human intellectual supremacy, functioning instead as conventional regulatory advocacy. Ball also noted that Olah's own remarks at the event directly contradicted the encyclical's assertion that AI cannot have subjective experiences. Observers quoted in the piece suggest the document is unlikely to significantly influence how technology companies develop AI or how governments regulate it.

Scotland's policy framework for attracting datacentre investment may be undermining its own climate commitments, according to an analysis by Edinburgh-based charity Action to Protect Rural Scotland. The country has no formal definition of what constitutes a "green datacentre," yet the designation carries real regulatory weight: developers who claim the label may receive favorable treatment from local planning authorities. One Edinburgh facility reportedly described itself as a green datacentre in planning submissions despite plans to install 200 diesel backup generators.

The underlying greenhouse gas analysis supporting Scotland's National Planning Framework 4 was conducted in 2022, before the release of ChatGPT, and has not been updated to account for AI's energy demands. More than a dozen datacentres are currently seeking planning permission in Scotland, with projected collective power use of roughly 6.2 gigawatts, about one and a half times Scotland's peak winter consumption. Over 100 datacentre projects across the UK have also requested gas connections due to grid constraints.

For legal professionals advising on energy, planning, or environmental compliance, the gap between policy language and measurable standards here illustrates the risks of undefined regulatory terms in fast-moving technology sectors.

Spotify CEO Alex Norström has publicly defended the company's recent move into AI-generated music, framing it as a responsible alternative to piracy and unregulated AI content. Last week, Spotify announced a premium feature allowing users to create AI-generated remixes and covers using music from participating artists, developed in partnership with Universal Music Group. The deal sent Spotify's shares up 16%.

Norström argues the feature gives artists a way to consent to AI use of their work and earn money from it. However, composer and copyright advocate Ed Newton-Rex cautioned that the feature's real risk depends on its design. If users can share AI remixes publicly, Newton-Rex warned, AI-generated content could flood the platform and crowd out human artists, creating pressure for even reluctant musicians to participate.

Key implementation details remain unresolved, including whether remixes will be shareable and how AI-generated content will be labeled. The announcement arrives amid broader legal and policy turbulence over AI and copyright, including ongoing lawsuits against Meta and OpenAI over training data, and two Meta employees individually sued last week for allegedly pirating a terabyte of books to train the Llama model.

UC Berkeley Law School is tightening its AI policy this summer, moving well beyond its 2023 rules, which faculty now consider too permissive. The updated policy prohibits students from using AI for conceptualizing, outlining, drafting, revising, editing, translating, or in any exam context. The 2023 policy had allowed brainstorming and conceptualization with AI tools.

Professor Chris Hoofnagle, who helped draft the new rules, frames the policy as a response to the rapid advancement of large language models. His concern is less about AI itself and more about what happens when students rely on it before mastering foundational legal skills. "If that lawyer cannot use their own analytical judgment to assess an AI output, that lawyer has very little value," he told Business Insider.

The policy comes with acknowledged limits. Hoofnagle notes that tools like Westlaw and Lexis now surface AI-generated summaries by default, making a clean line between permitted and prohibited use difficult to draw. Berkeley has also seen an increase in academic misconduct cases and has shifted more exams to in-person, proctored formats as a result.

A new report from Gallup and the Walton Family Foundation finds that while K-12 teachers are already using AI on the job, roughly eight in ten report receiving no formal guidance from school leaders on how to apply the tools. Gaps extend across specific use cases: around 71% received no guidance on using AI for feedback or coaching on their teaching, and 69% received none on one-on-one instruction or tutoring. Teachers in higher-needs schools were less likely to receive guidance than those in wealthier institutions.

Education experts and school administrators quoted in the report warn that the absence of structured support is pushing teachers to integrate AI in piecemeal ways that may not improve student outcomes. The CEO of All4Ed notes that the question is no longer whether students will use AI, but whether educators have the professional development and institutional backing to use it confidently. One ed-tech founder cautions that AI "can amplify the best parts of teaching and the worst," underscoring that volume of adoption matters less than strategic implementation.