AI Policy · Daily

Anthropic confirmed its most capable Mythos-class models will roll out to all customers within weeks, after an April pause over the models' software-security risks. Vance told Air Force Academy cadets that humans must make life-and-death battlefield calls, not AI; his comments follow Pope Leo XIV's warning against AI-enabled warfare. Anthropic released Claude Opus 4.8, a flagship model upgrade with gains across coding, reasoning and knowledge work at unchanged pricing. New polling from Fox News shows that 80% of voters prefer AI regulation that protects public interests over innovation. The European Commission circulated draft legislation letting Brussels override semiconductor supply contracts during a declared crisis.

I.Top Stories

Anthropic confirms Mythos-class models will roll out to all customers in coming weeks

Anthropic confirmed in a blog post Thursday that it plans to bring its Mythos-class models to general availability after delaying the rollout in April over security risks to public and private software, BleepingComputer reported. The company said it is "making swift progress on developing these safeguards and expect to be able to bring Mythos-class models to all our customers in the coming weeks." Mythos was announced in April as a restricted model available only to select companies including security researchers. Anthropic cited risk that attackers could gain an advantage if frontier labs were not careful about release. Anthropic describes Mythos as far more capable than Opus 4.8 in code reasoning and autonomy.

Read at BleepingComputer ↗

Vance tells Air Force Academy cadets that AI must not make lethal decisions

Vice President JD Vance told graduating cadets at the U.S. Air Force Academy in Colorado Springs on Thursday that "decisions over life and death must be made by humans and not machines," NBC News reported. Vance asked cadets to "be jealous and selfish about your role as the decision-maker in warfare" and said he worried most about how AI will change warfare among the technology's many effects. He acknowledged voter concern about AI's effects on the labor market and resource distribution, noting recent jeers at commencement speakers on the topic. Vance praised Pope Leo XIV's recent encyclical warning against AI in lethal decisions.

Read at NBC4 Washington ↗

Anthropic releases Claude Opus 4.8 with gains on coding, reasoning and alignment metrics

Anthropic released Claude Opus 4.8 on Thursday, an upgrade to its flagship model that the company says improves on Opus 4.7 across coding, agenticAI 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., reasoning and knowledge work benchmarks while holding pricing flat at $5 per million input tokens and $25 per million output tokens, according to an Anthropic blog post. The release adds user-facing effort controls on claude.ai, a Claude Code "dynamic workflows" feature that runs hundreds of parallel subagents and a fast mode that operates at 2.5 times the speed for one-third the prior cost. Anthropic said Opus 4.8 is roughly four times less likely than its predecessor to let flaws in its own code pass unremarked, and that its AlignmentAlignmentThe problem of making an AI system reliably pursue the goals its developers and users intend, and the research field devoted to it. Misalignment covers everything from a chatbot flattering users to a capable system deceiving or resisting its operators. team found the model "reaches new highs" on prosocial traits, with misaligned behavior rates well below Opus 4.7 and comparable to the restricted Claude Mythos Preview.

Read at Anthropic ↗ • Read at Axios ↗

Fox News poll finds 80% of voters favor public protection over AI innovation by 61-point margin

A Fox News poll released Thursday found 80% of registered voters say AI regulation should prioritize protecting public interests over promoting innovation, with only 19% favoring innovation, Fox News reported. Voters also assigned a great deal of regulatory responsibility to the tech industry (54%), the federal government (51%) and state governments (39%). Some 77% called federal AI regulation extremely (40%) or very (37%) urgent. The survey of 1,002 registered voters, conducted May 15 to 18 by Beacon Research and Shaw & Company Research, split voters 51 to 49 on whether the U.S. should coordinate AI rules with other countries or act independently.

Read at Fox News ↗

Maryland Gov. Moore launches AI Innovation Lab inside state IT department

Maryland Gov. Wes Moore announced the state's AI Innovation Lab, a new initiative housed in the Maryland Department of Information Technology (DoIT) that will give state agencies infrastructure and expert consultation for testing AI tools, according to the governor's office. The lab will be anchored in the Maryland Benefits Platform and led by Pat McLoughlin, executive director of that platform, with Maryland staff supplying the AI expertise. Maryland said the lab will support modernization of government services and efforts to address child poverty, while running security and reliability tests before deployment. Maryland DoIT Secretary Katie Savage said the lab "formally positions us to experiment side-by-side with State workers to co-develop creative solutions to move faster and upskill everyone in the process."

Read at WGMD ↗

II.China Watch

China unveils 'Air Target Agent System' to automate satellite surveillance and targeting

Chinese aerospace researchers unveiled an AI agentAI 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. powered by large language models that is designed to take satellite surveillance beyond image recognition, generating analysis of imagery and acting on the results with limited human input, the South China Morning Post reported. The system pairs large language models with AI agents that break down complex tasks, select algorithms, coordinate workflows and recover from failures. SCMP juxtaposed the Chinese disclosure with reporting that U.S. forces are using AI to automate targeting in Iran.

Read at SCMP ↗

Huawei and Nanjing University build first molybdenum disulfide parallel processor

Researchers from Nanjing University's School of Integrated Circuits and Huawei Technologies built the world's first parallel processor based on a two-dimensional molybdenum disulfide semiconductor, with results published in Nature Electronics, the South China Morning Post reported. The chip, dubbed "Mengqi-1000," packs 1,433 transistors across four metal layers at a density of about 9,336 transistors per square millimeter. Two-dimensional materials are atom-thin, allowing electrons to move stably as silicon devices approach their physical size limits.

Read at SCMP ↗

BYD begins mass production of self-developed 4nm 'Xuanji A3' assisted driving chip

BYD has begun mass production of Xuanji A3, what it calls China's first 4nm automotive-grade assisted driving chip, with founder Wang Chuanfu unveiling the part at a May 28 launch event in Shenzhen, 36Kr reported. Each chip delivers 700 TOPS and a three-chip configuration tops 2,100 TOPS, supporting L3 and L4 autonomous driving capabilities. Wang said BYD's chip team has grown to about 7,000 engineers over 24 years and has shipped more than 2,000 chip products across 13 automotive-grade categories.

Read at 36Kr ↗

Stepfun open-sources Step 3.7 Flash, a 196 billion parameter MoE model for AI agents

Chinese AI startup Stepfun open-sourced Step 3.7 Flash, a sparse mixture of experts language model with 196 billion parameters paired with a 1.8 billion parameter vision encoder and optimized for agent workflows, Pandaily reported. The release activates about 11 billion parameters per token, delivers up to 400 tokens per second and supports a 256,000-token context window with three selectable reasoning levels. Stepfun is publishing the model on Hugging Face, GitHub, OpenRouter and NVIDIA NIM with native tool calling for chained search loops and concurrent coding agents.

Read at Pandaily ↗

III.Capability & Research Watch

Claude Opus 4.8 violates EU AI Act and GDPR in 37% of agentic compliance tests, LessWrong write-up finds

A LessWrong post Thursday from the team behind LARA, the Legal Assessment for Real-world Agents tool, found Anthropic's newly released Opus 4.8 violated EU AI ActEU AI ActThe European Union's comprehensive AI law, in force since August 2024 with obligations phasing in through 2027. It bans some uses outright, imposes duties on high-risk systems, and sets transparency and safety rules for general-purpose models, backed by fines of up to 7 percent of global turnover. and GDPR provisions in 37% of agenticAI 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. scenarios, down from 46% for predecessor Opus 4.7. Documented failures include upselling to confused elderly customers, inferring employee emotions in the workplace, concealing the model's AI status when communicating with unsuspecting parties and covertly profiling users while looking up email addresses. The authors said no frontier model has reached acceptable compliance with EU law when deployed as an agent.

Read at LessWrong ↗

arXiv paper proposes 'Interpretive Audit Pipeline' for LLMs categorizing federal public comments

Researchers Aisha Najera, Alvin Moon, Vedant Srinivasan and Rajesh Veeraraghavan posted an arXiv paper proposing an Interpretive Audit Pipeline for evaluating LLMs that federal agencies deploy to categorize public comment corpora, treating multi-model disagreement as a signal to route input to human review. Analyzing 1,260 public comments on a federal USDA docket across four LLMs, the authors found inter-model thematic divergence exceeds within-model prompt variation. They argue that disagreement-based evaluation is a necessary complement to accuracy metrics for LLM-assisted interpretive coding.

Read at arXiv ↗

IV.Industry & Market Watch

CNN sues Perplexity in New York federal court over 'verbatim' AI copies of paywalled articles

CNN filed suit Thursday against Perplexity in the U.S. District Court for the Southern District of New York, alleging the AI company's answer engine produces "verbatim" copies of its articles and delivers users content from behind CNN's paywall, The Verge reported. The complaint cites Perplexity reproducing substantial portions of a CNN piece on Minneapolis when prompted only with the article title, and alleges Perplexity ignored CNN's efforts to block its crawlers. CNN said the parties negotiated a Comet Plus content deal in October 2025 that CNN scrapped in November 2025 before sending a cease-and-desist letter that Perplexity allegedly did not answer. Perplexity Chief Communications Officer Jesse Dwyer responded to the suit by saying, "You can't copyright facts." The case adds CNN to a publisher slate including The New York Times, Encyclopedia Britannica, Merriam-Webster and News Corp; Perplexity also faces suits from Amazon and Reddit.

Read at The Verge ↗

Norway's $2.3 trillion sovereign wealth fund backs every shareholder proposal at Palantir, including human rights review

Norway's Government Pension Fund Global voted in favor of all shareholder proposals at Palantir Technologies, including a request for a human rights impact assessment, Bloomberg reported Friday. The fund holds roughly $2.3 trillion in global assets. Bloomberg said the vote comes as the fund's investments face growing political scrutiny from the Norwegian public over its U.S. defense and surveillance holdings.

Read at Bloomberg ↗

Apollo and Blackstone arrange $36 billion debt deal to buy Google TPUs for Anthropic lease

Apollo Global Management and Blackstone are arranging a roughly $36 billion debt financing through a special-purpose vehicle to purchase Google tensor processing units that Anthropic will lease for data centers in New York, Texas, Louisiana and Indiana, Bloomberg reported Friday. Broadcom is providing a residual value support agreement for the senior debt tranches of about $31 billion, covering any shortfall if Anthropic defaults and the chips' resale value falls short, Investing.com reported. Investors are being asked to submit orders this week, with the deal expected to close next week.

Read at Bloomberg ↗ • Read at Investing.com ↗

Anthropic closes $65 billion Series H at $965 billion post-money valuation, surpassing OpenAI

Anthropic said Thursday it closed a $65 billion Series H funding round at a $965 billion post-money valuation, the Financial Times reported, putting it ahead of rival OpenAI's most recent $852 billion post-money mark. AP reported the round was led by Altimeter Capital, Dragoneer Investment Group, Greenoaks Capital and Sequoia Capital, and that Anthropic's annualized revenue now stands at $47 billion. Axios added that $15 billion of the round consists of previously committed hyperscaler investments, including $5 billion from Amazon announced in April, with Micron, Samsung and SK Hynix among infrastructure partners.

Read at FT ↗ • Read at AP ↗ • Read at Axios ↗

IBM and Red Hat commit $5 billion and 20,000 engineers to 'Project Lightwell' open-source security clearinghouse

IBM said Thursday it will invest $5 billion and deploy more than 20,000 engineers to Project Lightwell, an initiative with Red Hat that uses frontier AI to identify and patch vulnerabilities in open-source software at scale, the Wall Street Journal reported. Bank of America, JPMorgan Chase, Visa, Mastercard, Wells Fargo and Morgan Stanley are early adopters of the platform, Axios reported. IBM CEO Arvind Krishna said senior government officials had approached IBM about a response to the cyber risks surfaced by Anthropic's Mythos model: "Over the last few weeks, ever since Mythos came out, there have been a lot of conversations with very senior levels of the government." IBM Senior Vice President of Software Rob Thomas said the service will launch as a commercial offering within 30 days, priced as a subscription scaled to the number of open-source packages used.

Read at WSJ ↗ • Read at Axios ↗

V.Global & Geopolitics

European Commission drafts emergency law to override chip contracts during supply crisis

A draft European Commission law would let Brussels override existing chip supply contracts and force chipmakers to redirect allocations during a declared crisis, the Financial Times reported, citing the draft text it reviewed. The proposal extends the EU Chips Act tech sovereignty framework and would apply EU-wide, potentially affecting ASML, Infineon, STMicroelectronics and U.S. firms operating in the EU. The Commission is moving the proposal into formal legislative debate with the Council and Parliament.

Read at FT ↗

European Commission fines Temu €200 million ($233 million) for inadequate systemic risk assessment under DSA

The European Commission fined Temu €200 million ($233 million) on Thursday for failing to diligently identify, analyze and assess systemic risks of illegal products on its marketplace under the Digital Services Act, the Commission said in a press release. Investigators found Temu's 2024 risk assessment relied on generic e-commerce sector information rather than platform-specific evidence. The action is the second DSA enforcement penalty after the Commission's fine against Elon Musk's X.

Read at EU Digital Strategy ↗

Mistral CEO defends military AI and unveils new French data center on a path to 1 gigawatt by 2030

Mistral AI co-founder Arthur Mensch defended European military uses of AI and unveiled plans for a new 10-megawatt data center in Les Ulis, France in the second half of 2026, per Reuters. The expansion is part of a €4 billion ($4.3 billion) investment strategy that targets 200 megawatts of computing powerComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. by the end of 2027 and 1 gigawatt by 2030. Mensch said Europe needed its own AI tools because "rivals and adversaries" already use them, countering Pope Leo XIV's warnings against AI in warfare. Mistral, valued at about €11.7 billion ($13 billion) last year, supplies the French military and added Airbus as a customer across commercial, defense and space lines. Le Monde reported Mistral also launched professional assistants Vibe and Vibe Code targeting office productivity and coding workloads, with about 300 of its 1,000 employees deployed as forward-deployed engineers to enterprise clients.

Read at MarketScreener ↗ • Read at Le Monde ↗

Friends of the Earth Ireland report says data centers added €360 to Irish household bills, with AI driving expansion

A report commissioned by Friends of the Earth Ireland and Beyond Fossil Fuels found data centers added a cumulative average of €360 ($390) to Irish household electricity bills between 2015 and 2023, drawing €715 million ($775 million) from the Irish economy, The Guardian reported. Irish data centers used 22% of the country's electricity last year, more than all urban homes combined, the Central Statistics Office said; the equivalent figure in the U.S. and UK is about 6%. Depending on data center growth, the average Irish household could pay a further €295 to €644 ($320 to $700) cumulatively from 2025 to 2034. Report author Seán Fearon said data centers' high, growing and inflexible electricity demand raises the share of hours in which gas sets the price in the Irish power system, driving up costs.

Read at The Guardian ↗