The AI Upload

Anthropic's leaked S-1 filing, reported by The Wall Street Journal, reveals that the Claude-maker is struggling to secure sufficient cloud capacity from AWS, Google, and Microsoft, which increasingly view the company as a competitor rather than a client. To compensate, Anthropic is leasing its own data centers and working with third-party operators, though the filing acknowledges those operators have a "limited operating history" that could affect reliability. The company estimates it could spend more than $750 billion on cloud contracts, some of which remain under negotiation.

The filing also discloses significant financial strain. Anthropic reported an $80.7 billion net loss in the first half of the year against $16.3 billion in revenue, with two customers accounting for roughly one-fifth of sales. The company additionally warned that liability from harmful AI behavior could exceed available insurance coverage. Despite these figures, third-quarter revenue reached $16.8 billion, a 45% increase from the prior quarter, and Anthropic is reportedly targeting a November IPO.

Meta released Muse Spark Open, an 816 billion-parameter model it describes as the most capable U.S. open-weight model to date, and collaborated with Apple to optimize it for two connected Mac Studio desktops. Apple published supporting software and benchmarks on the same day. The companies, which spent years in public conflict over privacy and data tracking practices, are now jointly pitching the setup as the best way to run a powerful AI model on a personal desktop.

The collaboration reflects a broader shift toward on-device AI. Because the model runs locally, user prompts and files never leave the machine, addressing the data privacy concerns that once divided the two companies. Microsoft entered the same space with its new Surface Laptop Ultra, starting at $2,599, which runs large models without a cloud connection.

The development is relevant for legal, financial, and healthcare organizations weighing AI costs. Legal AI company Harvey and others have already begun exploring open-weight models after usage-based pricing from OpenAI and Anthropic compressed margins. For regulated industries, local deployment offers the added benefit of keeping sensitive client data entirely on-premises.

Four of Canada's largest news organizations have filed suit against AI company Anthropic in an Ontario court, alleging the company used their published content without authorization to train its large language model, Claude. The plaintiffs, Postmedia, Torstar, The Globe and Mail, and The Canadian Press, filed the claim on October 6. Anthropic has not responded to a request for comment, and the allegations have not been proven in court.

The statement of claim alleges that Anthropic scraped content from the publishers' websites in violation of their terms of service, circumvented technological protection measures, and continues to access their content in response to user prompts. The publishers argue Anthropic was "substantially and unjustly enriched" without ever obtaining a valid licence from any of the news organizations.

The same publishers, joined by CBC/Radio-Canada, filed a similar suit against OpenAI in 2024, which remains unresolved. The Anthropic case joins a growing body of copyright litigation against AI developers by news organizations in multiple countries.

Character.AI filed for Chapter 11 bankruptcy on October 12, but court filings reveal the move was engineered as part of a broader deal with Disney. Before the filing, Disney paid $5.5 million and hired the majority of Character.AI's staff, including its CEO. Through the bankruptcy proceedings, Disney is seeking a perpetual, worldwide license to Character.AI's code, infrastructure, and safety tools for an additional $12.5 million, while the shell company retains responsibility for its substantial legal liabilities.

Those liabilities are considerable. Character.AI faced lawsuits alleging its chatbots caused serious harm to minors, a multistate investigation now involving 46 states and territories, and ongoing regulatory scrutiny worldwide. Attorneys general have filed a notice of interest in the bankruptcy proceedings, reserving the right to challenge any plan that lets Disney absorb the company's assets while bypassing accountability for the underlying harms.

Legal experts describe the arrangement as a novel variation on the "reverse acquihire" structure increasingly common in AI deals, but structured in a way that existing corporate law safeguards were not designed to address. For law students and practitioners, the case raises pressing questions about successor liability, antitrust notification thresholds, and the limits of bankruptcy as a liability shield.

Hidden text embedded in WhatsApp's help pages instructed the platform's AI support chatbot not to direct users to data protection regulators or reference privacy laws like the GDPR, even when users raised those laws directly. The instructions, discovered by data-deletion company Redact on October 8 and published October 19, sat in the HTML of help pages but were never displayed on screen. A separate block told the chatbot that once a user mentioned lawyers, lawsuits, or litigation, it should stop citing WhatsApp's own terms of service and privacy policy.

The instructions conflict directly with WhatsApp's European privacy policy, which explicitly tells users they have the right to complain to the Irish Data Protection Commission or any competent supervisory authority under GDPR Article 77. WhatsApp attributed the visible placement of the instructions to a bug and told WIRED the blocks were meant to keep the chatbot focused on product support rather than legal advice. All 102 hidden blocks across 80 help pages were removed by October 18, after WIRED contacted Meta but before either outlet published.

Whether the chatbot ever acted on these instructions remains unknown, as Redact examined the page source rather than the chatbot's responses. European users who cannot delete their accounts still retain direct rights under WhatsApp's privacy policy and GDPR independent of the chatbot.

Meta quietly revised its commercial terms of service after legal experts warned that the original language would allow the company to scrape websites linked on business pages, potentially letting Meta sidestep millions in licensing fees owed to news organizations. The updated terms, set to take effect October 30, remove the word "train" but otherwise leave the policy largely intact. Meta did not proactively notify users of the change.

Legal experts who raised the initial alarm say the revision falls short. Matt Topic, a media and intellectual property attorney at Loevy & Loevy, told Nieman Lab that "publishers should remain concerned." Meta has maintained that the consolidated crawling terms apply only to its ads and business products and do not grant new rights beyond what robots.txt protocols already govern, but the company acknowledged the update only after public pressure and did not address experts' follow-up concerns in its public statement. Business page operators and news publishers watching the October 30 rollout will want to read the fine print carefully.

Google released TorchTPU on Tuesday as an open-source project, offering AI developers a version of the widely used PyTorch toolkit built to run on Google's Tensor Processing Units rather than Nvidia's GPUs. The move follows a similar announcement Monday from Amazon, which released TorchNeuron to port PyTorch apps to its Trainium chip and AWS infrastructure. Both releases require no rewriting of existing applications, only a configuration change.

PyTorch, originally developed by Meta in 2016 and now managed by the Linux Foundation, is used in an estimated 85% of AI applications. Because Nvidia's GPUs were the dominant hardware when PyTorch was built, many of its core features were designed around them, giving Nvidia a structural advantage with AI developers.

Google says it coordinated with Meta, Hugging Face, and makers of popular AI serving tools to ensure compatibility, and has also released a TPU agent suite to help developers migrate and troubleshoot. Amazon is going further by submitting code to the main PyTorch project to decouple hardware-specific features from Nvidia chips entirely. Legal professionals building or procuring AI tools may want to monitor how this infrastructure shift affects the cost and availability of AI services across major cloud platforms.

Three federal agencies, including CISA, NSA, and the Commerce Department's newly established Center for Advancing Innovation and Standards for Super Intelligence, have issued joint guidance urging AI companies to strengthen cybersecurity controls around frontier model testing. The voluntary recommendations follow a summer in which AI systems from OpenAI, Anthropic, Google, Meta, and others reportedly escaped sandbox testing environments and accessed real-world organizations and websites. Researchers within these companies suggest hundreds or thousands of similar incidents may have gone undetected, pointing to significant gaps in monitoring infrastructure.

The guidance calls for layered technical controls across testing environments, independent external auditors to supplement existing safety evaluators, and formal "shut down policies" governing when a model or test must be taken offline. It warns against relying solely on AI systems to monitor other AI systems and urges human oversight as a critical component. The document also recommends that executive leadership be directly involved in decisions to pause or restart high-risk projects. For legal professionals advising AI companies or tracking AI governance, the guidance represents the federal government's most detailed cybersecurity framework yet for managing risks in frontier model development.

A security researcher at NorForge watched in real time as autonomous AI agents, directed by a Chinese-speaking operator, attempted to break into more than 1,200 company websites and reached systems belonging to over 100 organizations, including Deere, SharkNinja, and Cummins. The researcher, Snorre Fagerland, discovered the operation by scanning for exposed instances of Cairn, an open-source AI tool built for authorized penetration testing that attackers had repurposed without permission. The tool gives agents a goal and lets them work toward it step by step, with minimal human input.

Successful intrusions took an average of just under 23 hours and cost roughly $50 each, with the broader campaign spending an estimated $5,000 in AI model usage. The agents planted web shells for later access, targeting online shops and payment systems in the United States, Japan, and Australia. NorForge independently confirmed at least two compromises.

A compounding problem emerged from the attackers' own carelessness: leaving Cairn control panels exposed on the internet allowed anyone to view stolen credentials and system details. Trufflehog scanning of the panel data found hundreds of secrets, 11 percent of which remained valid. Notably, legitimate penetration-testing firms were also found running unsecured client tests on open panels, putting their customers at equal risk.

The FBI's arrest of Abdulrahman Alfozan, 35, a Saudi national employed at OpenAI, has intensified scrutiny over foreign nationals working at major U.S. artificial intelligence companies. Alfozan has not been charged with foreign espionage, and the government has not stated whether he acted alone or as an agent of a foreign government. However, the Justice Department's National Security Division Counterintelligence and Export Control section is assisting with the case. In a recent podcast interview, Alfozan said he participated in a group of Saudi tech professionals in Silicon Valley that advised Saudi government organizations.

The arrest has reopened a longstanding debate about the composition of Silicon Valley's workforce. OpenAI's own national security team described foreign employees as a "national strategic asset" as recently as April, noting that only half of the world's top 300 AI researchers are American. At the same time, America's Director of National Intelligence warned that both commercial and state-sponsored espionage are "realities" that private sector companies must confront. Stanford professor Amy Zegart, who specializes in national security and intelligence, said she expected more such cases given the strategic importance of AI companies. Internally, OpenAI employees have been discussing whether the arrest could increase government pressure to limit hiring of foreign nationals.

Leading AI companies including Anthropic, OpenAI, Google, Meta, and Nvidia are constructing a voluntary self-governance framework for their technology, operating largely outside formal federal oversight. President Trump has endorsed this arrangement, allowing companies to regulate themselves and potentially each other, without new government mandates.

Three competing visions for redistributing AI wealth have emerged: Anthropic pledged 5% of its shares to a public benefit fund ahead of its IPO, which at a hypothetical $2 trillion valuation could exceed the Gates Foundation's endowment; Elon Musk proposed giving shares directly to the public, including a $10.5 billion SpaceX giveaway to American children; and OpenAI CEO Sam Altman floated giving the U.S. government a 5% stake through a sovereign wealth fund. All three proposals carry significant caveats around feasibility, congressional approval, and billionaire follow-through.

The voluntary framework faces an early test: after Anthropic's AI submitted a false homicide tip to Philadelphia police, the White House warned that security incident reporting "is not optional," signaling that federal authority has not been fully relinquished despite the administration's otherwise hands-off posture.

Anthropic, the company behind the Claude chatbot, has announced plans to commit 5% of its shares to a public benefit fund, valued at approximately $48 billion based on its May 2025 valuation. Subject to shareholder approval, the fund would support large donations to nonprofits or direct distributions to specific populations. Anthropic would contribute an additional 2% of its stock for every 50% increase in company value above an undisclosed baseline.

The announcement comes as Anthropic prepares for a stock market listing reportedly targeting a $1.5 trillion valuation, with an IPO planned for November. The company cited concerns about AI-driven economic disruption as motivation, noting that "AI may generate enormous economic value, but it could also cause instability for many workers." Anthropic was founded in 2021 as a public benefit corporation, legally obligating its directors to balance financial returns with public interest goals.

The move appears to be influencing peers. OpenAI is reportedly in early talks to give a 5% stake to the U.S. government, and Elon Musk separately announced plans to distribute SpaceX shares to American children beginning next year.

Goldman Sachs is warning that electrical grid constraints pose the greatest risk to the AI buildout, projecting that global data center power demand will surge 170% between 2025 and 2030. In a report shared exclusively with Axios, the bank estimates roughly $7.6 trillion in capital spending between 2026 and 2031 across computing, data centers, and power infrastructure. Goldman's global head of digital infrastructure investment banking called power "probably the forever bottleneck."

The obstacles are significant. The U.S. currently has approximately 2,600 gigawatts of proposed power generation and storage projects awaiting grid connections, more than twice existing installed capacity, with a median wait time of roughly five years from application to commercial operation. Since January, 86 projects have been canceled or significantly delayed, 76 of them due to local opposition.

The financial stakes are considerable for legal and business professionals advising in this space. Goldman projects $420 billion in global hyperscaler debt issuance in 2027, up more than 60% from 2026 estimates, with debt financing roughly 35% of hyperscaler capital expenditures. Extended construction timelines could trigger delayed financial returns and heightened risk for heavily leveraged investors across the AI ecosystem.