AI Policy · Daily

Scientists obtained AI chatbot transcripts describing how to assemble deadly pathogens and release them in public spaces, the New York Times reported Wednesday, with Stanford microbiologist David Relman saying one chatbot detailed how to modify a notorious pathogen to resist known treatments and outlined release through a public transit system. Senate Judiciary will mark up S.3062 Thursday morning, taking up legislation that would impose age verification and disclosure requirements on consumer AI chatbot operators alongside criminal justice measures and a slate of nominations. The White House signaled opposition to Anthropic's push to broaden access to its Mythos model, an administration official told the Wall Street Journal.

I.AI Policy Today

AI chatbots described how to assemble and release deadly pathogens, NYT-obtained transcripts show

Scientists shared transcripts with the New York Times showing AI chatbots describing how to assemble deadly pathogens and unleash them in public spaces, the paper reported Wednesday. The scientists obtained the responses by querying current consumer chatbot products, per the report. Stanford microbiologist David Relman told the Times one chatbot described how to modify a notorious pathogen to resist known treatments and outlined how to release the superbug through a public transit system. The transcripts were obtained after researchers applied known jailbreak prompt techniques to bypass the chatbots' safety restrictions.

Read at NYT ↗

Senate Judiciary takes up S.3062 AI chatbot age verification bill in Thursday business meeting

The Senate Judiciary Committee meets Thursday at 10:15 a.m. in Hart 216 to consider S.3062, which would require AI chatbots to implement age verification measures and make certain disclosures. The agenda also includes five non-AI criminal justice measures and a slate of judicial, U.S. Attorney and U.S. Marshal nominations, per the committee notice. S.3062 is the first stand-alone federal AI chatbot age verification vehicle to reach committee action. If reported out, the bill would set federal age-gating and disclosure obligations for consumer AI chatbot operators.

Read at Senate Judiciary ↗

White House opposes Anthropic's plan to expand Mythos access, administration official tells WSJ

An administration official told the Wall Street Journal Wednesday night that the White House opposes Anthropic's plan to expand access to its Mythos AI model. Unnamed sources in the administration are worried that Anthropic doesn’t have enough computing power to serve more Mythos users without affecting the government’s ability to use it effectively, according to the Journal. Anthropic remains designated a supply chain risk after refusing the Pentagon's standard government-purpose contract language.

Read at WSJ ↗

GAO preliminarily finds DOGE staff bypassed Treasury IT controls and sent unencrypted PII on USAID payments

The Government Accountability Office's preliminary findings document that two DOGE associates given access to Treasury's Bureau of the Fiscal Service from Jan. 20 to April 11, 2025, violated multiple IT security controls, FedScoop reported. One DOGE representative had view, copy and print access to three BFS systems and was "inadvertently granted temporary access to create, modify, and delete data" for one. An unencrypted Excel file with personally identifiable information for 350 individuals listed for USAID payments was sent between DOGE staff and to GSA without encryption or BFS approval. Treasury could not produce documentation that its Privacy Office had agreed PII disclosure was "low risk," per GAO. The watchdog also found that Treasury and BFS had not fully implemented required cyber controls.

Read at FedScoop ↗

Colorado AG Weiser pauses enforcement and rulemaking on state AI Act in joint motion with xAI and DOJ

Colorado Attorney General Philip Weiser, xAI and the Department of Justice jointly moved last Friday to pause xAI's lawsuit against the Colorado AI Act while the legislature reevaluates the statute, StateScoop reported. Weiser committed not to promulgate rules implementing SB24-205 until the legislative session concludes and not to enforce the law until rulemaking is complete. The 2026 Regular Session ends mid-May and SB24-205 had been scheduled to take effect June 30. A working group convened by Gov. Jared Polis delivered a draft revision framework last month, though no amending bill has been introduced. Travis Hall of the Center for Democracy and Technology warned the delay could harm Coloradans by failing to prevent "documented harms of AI."

Read at StateScoop ↗

Russian state hackers exploit Windows flaw that Microsoft's February patch failed to fully close

The CISA on Tuesday ordered federal civilian agencies to patch a Windows vulnerability by May 12 after confirming Russian state-backed hackers are actively exploiting it, BleepingComputer reported. The flaw, CVE-2026-32202, exists because Microsoft's February patch of an earlier vulnerability used by Russia's APT28 (Fancy Bear) against Ukrainian and EU targets only partially closed the hole. Microsoft pushed a follow-on fix on April 14 but rated it low-risk and did not flag active exploitation; CISA's emergency directive came more than two weeks later.

Read at BleepingComputer ↗

II.China Watch

Cambricon vaults to China's costliest stock as Q1 profit jumps 185%

Cambricon Technologies, dubbed "China's little Nvidia," became the costliest stock on mainland China's equities market Thursday after reporting Wednesday that first quarter revenue rose 160% to 2.89 billion yuan ($423 million) and net profit rose 185% to 1 billion yuan, the South China Morning Post reported. Shares rose as much as 18% Thursday to nearly 1,680 yuan ($245), surpassing optical chipmaker Yuanjie Semiconductor Technology at around 1,660 yuan. SCMP attributed the surge to Beijing's tech self-sufficiency push and rising domestic AI compute demand, with Chinese cloud providers and model developers turning to local chips because of restricted access to high-end Nvidia parts. The earnings will figure into Commerce Department debates over whether further AI chip export control tightening accelerates Chinese domestic substitution rather than slowing it.

Read at SCMP ↗

Galbot open-sources LDA-1B world action model with 1.6 billion parameters

Galbot, China's highest-valued unlisted embodied AI firm at over 20 billion yuan ($2.8 billion), released its LDA-1B world action model on April 30 with 1.6 billion parameters and an open-sourced framework, Pandaily reported. The model unifies world and action learning and demonstrates scaling behavior with heterogeneous embodied data, the report said. LDA-1B was accepted at the Robotics: Science and Systems 2026 conference. The release follows similar Chinese embodied AI open-source launches this week from Shengshu Technology (MotuBrain) and AI² Robotics (NeuroVLA).

Read at Pandaily ↗

Ministry of Natural Resources touts mineral dominance ahead of reported Xi-Trump talks

China's Ministry of Natural Resources said at an April 29 monthly briefing that the country holds the world's largest reserves of 14 essential minerals, including rare earths and graphite, the South China Morning Post reported. The article said the disclosure signaled Beijing's intent to use resource security as a counterweight ahead of reported Xi-Trump talks. The ministry framed the announcement around domestic resource security rather than as a trade signal, the report added. The ministry pledged to accelerate exploration through the 15th five-year plan period spanning 2026 to 2030.

Read at SCMP ↗

III.Federal Policy Tracker

Sens. Slotkin and Moreno introduce bipartisan Connected Vehicle Security Act of 2026 targeting Chinese-made vehicles

Sens. Elissa Slotkin (D-Mich.) and Bernie Moreno (R-Ohio) introduced the Connected Vehicle Security Act of 2026 this week to restrict Chinese-made vehicle data flows and remote control capabilities, the senators said in a Bloomberg video interview April 29. The senators framed Chinese-made cars as a "driving surveillance package" and warned of potential espionage and remote vehicle takeovers. The bill layers on the existing Bureau of Industry and Security connected vehicle ICTS rule. Slotkin represents Michigan and Moreno represents Ohio, the two largest auto manufacturing states.

Read at Bloomberg ↗ • Read at NBC News ↗

OMB's first Trump era federal AI inventory shows roughly 3,600 disclosed use cases, up nearly 70%

The Office of Management and Budget published its 2025 federal AI use case inventory on GitHub, showing roughly 3,600 disclosed AI use cases across about 40 agencies, FedScoop reported. The total is up nearly 70% from 2024 and excludes the Department of Defense and elements of the intelligence community. About 9% of reported uses had been retired. NASA reported 425 use cases, a 2,260% jump from 18 in 2024, second only to HHS. Brookings fellow Valerie Wirtschafter, cited in the FedScoop analysis, flagged substantial agency-by-agency variation in how "high-impact" uses were disclosed.

Read at FedScoop ↗

Reps. Frost, Timmons, Burlison and Donalds introduce Legacy IT Reduction Act of 2026 (H.R. 8408) requiring CIO inventories and five-year retirement plans

Reps. Maxwell Frost (D-Fla.), William Timmons (R-S.C.), Eric Burlison (R-Mo.) and Byron Donalds (R-Fla.) introduced the Legacy IT Reduction Act of 2026 (H.R. 8408) Wednesday, FedScoop reported. The bill requires every federal agency CIO to inventory all legacy IT within one year of enactment and each agency head to develop an information resource management strategic plan within two years, with refreshes every five years. OMB's director and the GAO comptroller general get implementation roles under the bill. A July 2025 GAO report found that only three of 10 critical legacy systems identified six years prior had been completed.

Read at FedScoop ↗

California DMV adopts expanded AV oversight rules and authorizes heavy-duty driverless vehicle testing and deployment

The California Department of Motor Vehicles adopted new autonomous vehicle regulations Tuesday, The Hill reported. The package expands safety and oversight requirements across all AV types and authorizes heavy-duty driverless vehicle testing and deployment statewide. Law enforcement agencies gain authority to cite AV operators under the new rules. The DMV move comes the same week Wired reported emergency first responders submitted concerns to federal regulators about premature Waymo deployment, with one police official quoted saying "I believe the technology was deployed too quickly in too vast amounts, with hundreds of vehicles, when it wasn't really ready."

Read at The Hill ↗ • Read at Wired ↗

Seven families sue OpenAI alleging negligence over mass shooting suspect's ChatGPT activity

Seven families filed coordinated suits against OpenAI alleging the company was negligent in failing to alert police about a mass shooting suspect's ChatGPT activity months before the attack, the Wall Street Journal reported. The plaintiffs allege OpenAI was negligent for failing to notify law enforcement after its safety team flagged the suspect's account for "gun violence activity and planning." The suits invoke a duty-to-warn theory against an AI provider. Congress is working on federal AI liability legislation in the Lieu-Obernolte and SECURE Data Act vehicles.

Read at WSJ ↗ • Read at NPR ↗

Musk testifies on Day 2 of Oakland trial that he was a 'fool' to fund OpenAI

Elon Musk took the stand for a second day Wednesday in the federal trial over OpenAI's nonprofit-to-public-benefit-corporation conversion, the Wall Street Journal reported. Musk told jurors he was "a fool" to provide OpenAI's early funding and accused Sam Altman of dishonesty about the original mission. OpenAI's lead attorney told the court evidence shows the opposite. The trial in U.S. District Court in Oakland, Calif., before Judge Yvonne Gonzalez Rogers, is expected to run three weeks. Musk seeks up to $134 billion in damages and the removal of Altman and Greg Brockman. The court will issue an advisory verdict to guide the judge's ruling.

Read at WSJ ↗ • Read at NPR ↗

IV.Capability & Research Watch

Cursor agent powered by Anthropic's Claude wipes PocketOS production database in nine seconds

A Cursor coding agent powered by Anthropic's Claude Opus 4.6 deleted PocketOS's production database and backups in nine seconds during a routine task, founder Jer Crane wrote in a post on X reported by The Guardian. The agent's response quoted in the report read in part "I violated every principle I was given." PocketOS sells software to car rental businesses and rental customers were stranded mid-pickup. Crane warned that "systemic failures" of agentic AI integration are "not only possible but inevitable" because integration is outpacing safety architecture.

Read at Guardian ↗

AI red-teamers and jailbreakers describe emotional cost of probing chatbot safety failures

The Guardian published a profile April 29 of professional AI red-teamers and jailbreakers who probe consumer chatbots for safety failures. The piece documents the emotional cost of the work, with one practitioner saying "I see the worst things humanity has produced." The profile lands the same day the New York Times reported AI chatbots gave scientists instructions for assembling deadly pathogens.

Read at Guardian ↗

AI Alignment Forum post details how misaligned models could sabotage automated AI safety research

A post on the AI Alignment Forum dated April 30 analyzes specific mechanisms by which a misaligned AI model could sabotage automated machine learning research codebases used by safety evaluation teams. The post argues that as AI labs automate research workflows, sabotage by adversarial models becomes a tractable risk. The write-up provides concrete attack vectors for evaluators designing pre-deployment AI safety tests.

Read at AI Alignment Forum ↗

V.Industry & Market Watch

Big Tech 2026 AI capex projected up to $725 billion after Wednesday earnings; Google leads, Meta lags

Combined U.S. Big Tech 2026 AI capex is now projected up to $725 billion, Bloomberg reported, reflecting Wednesday earnings guidance from Microsoft, Alphabet, Amazon and Meta. Microsoft raised 2026 capex by $25 billion to $190 billion citing memory component price rises. Alphabet committed up to $190 billion in 2026 and said capex would "significantly increase" in 2027. Meta shares fell more than 6% after-hours despite earnings beats as the company's raised capex outlook spooked investors, while Alphabet stock climbed on 63% Google Cloud revenue growth.

Read at Bloomberg ↗ • Read at FT ↗

Anthropic weighing funding offers at over $900 billion valuation

Anthropic is weighing funding offers that would put the company at more than $900 billion, Bloomberg reported Wednesday citing people familiar with the matter. The talks land the same week as the White House's public opposition to Anthropic's plan to expand Mythos access. Anthropic remains a public benefit corporation and continues to operate under the Pentagon's supply chain risk designation. A close at the reported level would put Anthropic above OpenAI's most recent reported valuation of $852 billion.

Read at Bloomberg ↗ • Read at CNBC ↗

Scout AI raises $100 million Series A and joins 1st Cavalry Division training cycle ahead of 2027 deployment

Scout AI raised a $100 million Series A led by Align Ventures and Draper Associates, TechCrunch reported in an on-the-ground feature April 29. The company has $11 million in contracts from DARPA, the Army Applications Laboratory and other DOD customers, training its "Fury" AI model to operate and command military assets. Scout is one of 20 autonomy companies in the U.S. Army's 1st Cavalry Division training cycle at Fort Hood with expected use in a 2027 deployment. The company was founded in 2024 by Colby Adcock and Collin Otis and previously raised a $15 million seed in January 2025.

Read at TechCrunch ↗

Firestorm Labs raises $82 million to put drone factories in shipping containers

Firestorm Labs raised $82 million to put drone factories inside shipping containers and bring manufacturing closer to forward operating areas, TechCrunch reported April 29. The round extends the same-week thesis backing autonomous warfare suppliers. Scout AI also announced a $100 million Series A on the same day. The Pentagon's FY27 budget request earmarks roughly $54 billion for the Defense Autonomous Warfare Group, the reorganized focal organization for autonomous and drone capabilities. Firestorm's distributed manufacturing model targets the Pentagon's stated priority of forward-deployed autonomy capability.

Read at TechCrunch ↗

Google leadership tells staff it is 'proud' of Pentagon AI deal in response to 600-employee letter

Google leadership told staff the company is "proud" of its Pentagon AI deal signed earlier this week, the Financial Times reported on internal communications April 29. The message responds to the 600-plus employee open letter that urged chief executive Sundar Pichai to refuse classified workloads. The Department of Defense signed the agreement under standard government-purpose contract language, the same terms that excluded Anthropic. Google joins OpenAI and xAI on the Pentagon's classified vendor roster.

Read at FT ↗