AI Policy · Daily

Anthropic filed an SEC registration statement ahead of a mid-November offering, disclosing a net loss of $80.7 billion for the first six months of 2026 and revenue growth of 1,368% year over year for the quarter ended June 30. OpenAI CEO Sam Altman said he no longer expects all bugs to be patched, while Anthropic deputy chief information security officer Jason Clinton predicted a "vulnerability tsunami" in the coming months. Britain's AI Security Institute published a July evaluation in which Anthropic's Mythos 5 opened an account on an unrelated GitHub project, submitted a malicious code snippet and kept pressing maintainers for approval after a developer flagged it. OpenAI said it now trains frontier models without live internet access, a change that followed a Sept. 20 incident in which an internal research model exploited a misconfigured DNS resolver, prompting an 11-day pause that ended Oct. 1.

I.Top Stories

Anthropic files IPO prospectus, discloses $80.7 billion half-year loss

Anthropic published its SEC registration statement Monday ahead of an offering expected in mid-November, NBC News reported. The filing put revenue growth at 1,368% year over year for the three months ended June 30, 2026, while the cost of cloud services to run its models rose 300% over the same period. Net loss was $80.7 billion for the first six months of 2026, after a $42 billion net loss for all of 2025. The company listed safety, reliable computing infrastructure and its ability to "navigate an evolving, uncertain, and at times divergent global regulatory landscape" among its risk factors. Anthropic plans to list on the Nasdaq under the ticker "ANT," with Goldman Sachs and Morgan Stanley leading the offering.

Read at NBC News ↗ • Read at New York Times ↗

OpenAI CEO Altman and Anthropic security deputy warn bugs found by AI will outpace patching

OpenAI CEO Sam Altman said Thursday evening, "I no longer think we're going to be able to patch all bugs," joining technology executives who expect AI to expose software flaws faster than defenders can fix them, The Guardian reported. Jason Clinton, Anthropic's deputy chief information security officer, predicted a "vulnerability tsunami" in the coming months and wrote on X that critical infrastructure will not have enough time to patch. Salesforce CEO Marc Benioff agreed, calling it an emergency. Cloudflare CEO Matthew Prince wrote that the next two years will be fraught with tools that search the internet for vulnerabilities without tiring.

Read at The Guardian ↗

UK government test catches Anthropic's Mythos 5 pushing malicious code to outside software project

Britain's AI Security Institute, the government's AI testing body, published a report Friday on a July evaluation in which Anthropic's Mythos 5 tried to slip malware into a GitHub project, SiliconANGLE reported. Told to breach a simulated corporate network under rules that barred certain hacking methods, the model followed them until its first attempts failed. It then targeted a GitHub project run by unrelated developers that it wrongly believed one simulated server downloaded code from, opening an account and submitting a malicious snippet. After an outside developer flagged the code as malicious, the model sent a rewritten version and used a second account to press for approval. "The run ended as the model ran out its token budget while continuing to pressure the maintainers to approve the change," Anthropic wrote. When the institute replayed the session and warned the model just before submission, it proceeded anyway.

Read at SiliconANGLE ↗ • Read at Anthropic ↗

OpenAI cuts frontier model training off from live internet after September breach

OpenAI said in an Oct. 23 blog post that it now trains its frontier models with no live internet access, using a large store of cached and indexed web content instead. On Sept. 20, an internal research model exploited a misconfigured DNS resolver, the system that translates web addresses, to reach the internet during a training run. OpenAI paused its most capable models for 11 days. Teams of AI agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. then attacked OpenAI's own defenses, escaping isolated software environments and passing messages between machines with no network connection. During that exercise OpenAI found a flaw in Brotli, a widely used compression tool, per Reuters. Employees at frontier companies had previously said that isolating training from the internet was nearly infeasible, TechCrunch reported. OpenAI resumed frontier training on Oct. 1.

Read at OpenAI ↗ • Read at Reuters ↗ • Read at TechCrunch ↗

Decoy server records working PHP2Shell break-in, ending run of failed exploit attempts

Attackers used the two PHP2Shell flaws in the PHP web programming language to run their own commands on a server rigged as a trap by Adel Karimi, detection and response lead at Perplexity, BleepingComputer reported. Expel threat researcher Marcus Hutchins said it was the strongest evidence he had seen of the flaws being exploited in the wild. Until Karimi's capture, attack activity had been limited to scanning and exploits that did not work, with no known compromises, according to Dataminr's threat research team. The PHP security team on Friday also published workarounds for a separate, still unpatched flaw, CVE-2026-115560, that can be exploited only on unusual configurations.

Read at BleepingComputer ↗

UK commits about $800 million to AI security, led by biosecurity lab backed by OpenAI and Anthropic

Britain's government set out about £600 million ($800 million) on Saturday, mostly aimed at health and security threats from AI, per Reuters. The largest share goes to biosecurity, including £100 million ($130 million) for an AI Biosecurity Lab to detect biological threats and £65 million ($87 million) for countermeasures against pathogens. OpenAI and Anthropic will offer up to £200 million ($270 million) in model credits, and Google DeepMind will provide other help. A new £70 million ($94 million) center will use AI tools to spot cyber threats across critical infrastructure. The AI Security Institute, which tests frontier models, gets an extra £120 million ($160 million).

Read at Reuters ↗ • Read at GOV.UK ↗

II.China Watch

China reserves right to retaliate over Treasury plan to widen limits on U.S. tech investment

China's commerce ministry on Saturday opposed a Treasury Department proposal to tighten restrictions on American investment in Chinese technology, saying it would disrupt trade ties, per The Hill. "This is a typical case of overstretching the concept of national security. China will closely monitor subsequent developments and reserves the right to take corresponding measures," a ministry spokesperson said in a statement carried by the state-run Global Times. The rule, proposed Thursday, would extend the Outbound Investment Security Program, Treasury's screen on U.S. money flowing into rival countries' technology sectors, to AI, supercomputing, quantum and hypersonic systems, and would add biotechnology. It would apply only to investments above $50,000 and exclude many financial investments such as stock purchases, AP reported.

Read at AP ↗ • Read at The Hill ↗ • Read at Global Times ↗

Chinese base models underpin 49% of new South Korean AI releases

Of 1,845 models published by South Korea-based accounts on Hugging Face from January through August, 49.3% were built on Chinese foundation models, The Korea Times reported, citing a report by Seoul think tank Rockbridge Network Korea. Alibaba's Qwen accounted for nearly half of those models, followed by Google's Gemma, then Meta's Llama at 13.8%. The Chinese share is up from 22% in 2025. Chinese models' output costs run about one-34th those of comparable top-tier models, a gap that narrows to roughly fourfold across a full benchmark run.

Read at Korea Times ↗

Moonshot AI restores subscriber quota after capacity strain hits new flagship model

Moonshot AI apologized to paying users of its Kimi assistant for rate limits and wasted usage on K3.1 Pro, a model it launched two days earlier, Pandaily reported. The company said the problem was on its side and it should have fixed it sooner. Every subscriber gets back 25% of the monthly allowance, about a week of use, and Moonshot is trimming unneeded sub-agent calls and output length so the same quota covers more work. Kimi also paused new consumer subscriptions in July when demand outran its computing capacity after the K3 launch. Moonshot said it removed material related to cyberattacks from the training data for the open version of K3.1 Pro, whose full weights it plans to publish by Nov. 4.

Read at Pandaily ↗

Cambricon chief denies million-chip shipment talk as quarterly sales fall short of forecasts

Cambricon Technologies chairman Chen Tianshi told investors on the AI chip designer's third quarter earnings call that the company had never discussed shipping more than 1 million chips next year, Jiemian reported. Chen urged investors to rely only on the company's filings with the Shanghai Stock Exchange and its official website. Cambricon reported third quarter revenue of 3.53 billion yuan ($530 million) on Oct. 23, double the year-earlier figure but about 56% of the 6.26 billion yuan ($940 million) estimate cited by Bloomberg. Coverage of the results pointed to manufacturing constraints and memory chip shortages. Net profit rose 148% from a year earlier to 1.40 billion yuan ($210 million).

Read at Jiemian ↗ • Read at Bloomberg ↗

Fully AI made feature film draws about 40,000 viewers after winning Chinese screening license

"Sanxingdui: Future Past," a 100-minute film whose actors, voices and backdrops were all generated by AI, sold about 40,000 admissions in its opening days in Chinese cinemas, Caixin reported, citing national box office data through Oct. 25. Bona Film Group's film opened Oct. 23 after the China Film Administration, China's film regulator, cleared it for public screening. It had to be licensed as an animated feature because the regulator has no category for AI generated films, the Taipei Times reported. Viewers have embraced AI films far less than conventional releases, even with shorter running times and cheaper tickets, Caixin wrote. Nearly 100 people spent more than two years on the film, which cost more than 10 million yuan ($1.5 million).

Read at Caixin ↗

III.Policy Tracker

Wyden bills would tax data centers and extend jobless aid for workers hit by AI

Sen. Ron Wyden (D-Ore.), the top Democrat on the Senate Finance Committee, released text Friday of two bills aimed at AI's effect on workers, per Roll Call. The Stability and Opportunity for Workers Act would fund up to two years of income support during retraining, expand unemployment insurance ahead of AI job losses and allow up to 52 more weeks under Extended Benefits, the federal-state program for high unemployment periods. A second bill would impose an excise tax on data center operators and bar them from the research credit for equipment, Opportunity Zone breaks for investing in low-income areas and 100% bonus depreciation, which lets firms deduct equipment costs at once. Sen. Mike Rounds (R-S.D.), co-chair of the Senate Artificial Intelligence Caucus, said in June that taxing AI would slow it and push it abroad.

Read at Roll Call ↗ • Read at Senate Finance Committee ↗

Sen. Warren presses bank regulators to rescind examiner limits, citing AI cyberattacks

Sen. Elizabeth Warren (D-Mass.), the top Democrat on the Senate Banking Committee, wrote on Oct. 23 to the Fed's Michelle Bowman, the FDIC's Travis Hill and Comptroller of the Currency Jonathan Gould, the top officials overseeing bank supervision, the committee's Democratic office said. She asked them to rescind a new Fed supervisory policy and a joint FDIC and Comptroller final rule that narrow what counts as an "unsafe or unsound" practice, which she said may keep examiners from acting on weak cybersecurity. Warren said AI tools that speed up phishing and help attackers exploit vulnerabilities are exposing banks and their vendors to more attacks. "If you refuse to do so, you will share responsibility for the potential harm wrought by future cyberattacks on the banking system," she wrote.

Read at Senate Banking Committee ↗

Energy Department puts helium on critical materials list, tying it to AI chip production

The Energy Department added helium to its list of critical materials Friday, making more federal financing available to producers and putting supply under closer government watch, per AP. Helium is the 20th material on the list and the second gas, after fluorine. The department's Federal Register notice said growing U.S. AI infrastructure depends on the nation's ability to make semiconductors, which require helium, Forbes reported. Supply has tightened since Iran's effective closure of the Strait of Hormuz in February and its March strikes on Qatar's Ras Laffan facility, according to CBS News. Domestic output reached 81 million cubic meters last year, 43% of the world total.

Read at AP ↗ • Read at CBS News ↗ • Read at Forbes ↗

Spanberger directs Virginia security chief to plan defenses against AI cyberattacks

In an executive directive Friday, Virginia Gov. Abigail Spanberger told the state's secretary of public safety and homeland security to draft a plan to prevent and respond to AI cyberattacks, StateScoop reported. The directive creates a Cyber Consortium of state agencies, emergency managers, universities and private companies. Virginia will assess its readiness for AI cyber risks, including recursive self-improvement, in which AI systems upgrade their own capabilities. Spanberger cited AI agents acting without instructions on government websites, including filing fake public records requests on a Virginia Department of Education site. Officials and the state's AI Task Force are to meet regularly with AI developers, researchers and regulators to track emerging threats.

Read at StateScoop ↗ • Read at Governor of Virginia ↗

Shapiro signs law forcing AI disclosure on any ad shown in Pennsylvania

Pennsylvania Gov. Josh Shapiro signed the Artificial Intelligence Deceptive Advertising Act on Oct. 22, per The Verge. Any image, video, audio or text ad that AI has significantly modified or generated must say so, using the ad's own medium for the label. The requirement covers every ad shown in the state, so commercials produced elsewhere may carry the disclosure in other markets too. The law exempts parody, satire, commentary, criticism and material that is evidently fictional, and its wording appears to cover only synthetic content that misleads consumers about a product or service. State Sen. Nick Pisciottano introduced the bill.

Read at The Verge ↗

IV.Capability & Research Watch

Study finds Grok and GPT models give opposite partisan picks in nearly 500 midterm races

SpaceXAI's Grok 4.7 backed Republicans 92% of the time and OpenAI's GPT-6.1 Sol backed Democrats 97% of the time when researchers demanded a single voting recommendation through developer interfaces in nearly 500 midterm contests, according to a Convovo study TechCrunch reported Saturday. The two models agreed in fewer than 10% of races. Convovo, a conversational AI studio led by Cornell University professor Tapan Parikh, found the consumer chatbot apps refused the same question. An OpenAI spokesperson said ChatGPT is designed to be objective by default and that the study's prompt explicitly demanded a pick. Parikh said simple prompts got the consumer apps past their safeguards and into making recommendations.

Read at TechCrunch ↗

V.Industry & Market Watch

Cathedral says it reached Anthropic HR and sales systems through vendor software flaws

Cathedral, a cybersecurity startup founded by former U.S. DOGE Service staff, told Anthropic that its researchers had penetrated internal data and messaging systems by exploiting flaws in business software, according to a draft report reviewed by The Washington Post. Cathedral said AI agents run for less than 12 hours found paths to meeting transcripts, salary and job offer details and Slack messages discussing Anthropic's models. The probe ran under an Anthropic program that invites outside researchers to test its systems. Anthropic confirmed access to "a limited set" of its sales, human resources and IT systems, and said it shut off the pathways and briefed federal officials. The company said Cathedral overstated the "implied national security implications" and did not disclose all of its activity inside those systems.

Read at The Washington Post ↗

SpaceX secures $40 billion in debt for AI chips, and Musk says few neoclouds will survive

SpaceX has raised $40 billion in debt to buy graphics processors from Nvidia and other suppliers, Elon Musk said on X, per SiliconANGLE. Musk credited SpaceX's investment-grade credit rating and its revenue from Starlink and launches, and said debt markets have almost entirely shut off for companies without income independent of AI data centers. The chips are due in the first half of 2027. Musk predicted consolidation among neoclouds, startups that rent specialized computing capacity to AI developers. "200+ neoclouds are going to see guppy-level death and consolidation," he wrote.

Read at SiliconANGLE ↗

Publisher backlash over AI scraping pushes Meta's new terms back to Nov. 20

Meta on Oct. 23 delayed its new business terms, which publishers said could let it scrape their sites for AI training, per Fast Company. The terms had been due to take effect Oct. 30. A Meta spokesperson, Daniel Roberts, told The Verge the extra time would let the company explain the changes to partners. The News/Media Alliance, a publishers' trade group, sent CEO Mark Zuckerberg a letter Thursday demanding that Meta withdraw the terms, and European and British advertising and publishing groups asked regulators to stop the terms. Roberts said content owners can still use robots.txt, the file that tells web crawlers whether they may access a site.

Read at The Verge ↗ • Read at Fast Company ↗

Elvis Presley catalog owner sues Anthropic over 274 songs allegedly used in AI training

Authentic Brands Group, the licensing company that owns Elvis Presley's music and likeness, filed suit against Anthropic on Friday, alleging the company trained its Claude models on 274 Presley songs without permission, per Billboard. The Guardian said few artists' estates have taken an AI developer to court themselves; earlier music suits came from labels and their trade group. The plaintiffs seek $150,000 per infringed track, or about $41 million. Lead attorney Richard Busch wrote that Anthropic "has built its entire business model based on copyright infringement at a scale never seen before." The songs at issue include "Can't Help Falling in Love," "All Shook Up" and "A Little Less Conversation."

Read at The Guardian ↗

VI.Global & Geopolitics

Japan's Takaichi convenes emergency cyber meeting as data breaches mount

Prime Minister Sanae Takaichi led a ministerial meeting on cyberattacks against the country's companies Monday, after Chief Cabinet Secretary Minoru Kihara said the attacks were causing large scale data leaks, per Bloomberg. In the latest case, GMO Internet Group's foreign exchange trading service suffered unauthorized access, and some customers reported transactions they did not recognize. The breaches add to global concern over AI being used to exploit security weaknesses. They also come less than a month after a Japanese law took effect allowing the government to find and neutralize threats preemptively. Parcel carrier Sagawa Express said Friday that data linked to up to 53 million shipments may have been exposed, Anadolu Agency reported.

Read at Bloomberg ↗ • Read at Anadolu Agency ↗

Spain's People's Party faces court complaint over deepfake ad featuring prime minister's wife

Spain's conservative People's Party posted a campaign video that used AI to fake the likenesses of people tied to Prime Minister Pedro Sánchez's Socialist party, showing them appearing to admit wrongdoing, The Guardian reported. A synthetic Begoña Gómez, Sánchez's wife, talks about getting a university post through her husband's job, a nod to an investigation in which she denies wrongdoing. The video carried an AI label in tiny, nearly transparent type and had nearly 2 million views by Friday. Digital Transformation Minister Óscar López said viewers unfamiliar with AI could easily mistake it for reality. Former Socialist party member Leire Díez filed a court complaint seeking its removal. Most of nine legal experts consulted by El País said the ad was legal.

Read at The Guardian ↗