AI Policy · Daily

Treasury Secretary Bessent and Fed Chair Powell convened an emergency AI cyber-risk summit with bank CEOs over cybersecurity threats from Anthropic's Mythos model, which generated working exploits for zero-day vulnerabilities 72.4% of the time. xAI sued to block Colorado's AI anti-discrimination law on First Amendment grounds, while Florida opened a criminal probe of OpenAI over ChatGPT's alleged connection to the Florida State University (FSU) shooting that killed two people. AI chip demand remains resilient, with TSMC posting a 35% year-over-year revenue jump amid the Iran conflict, and Amazon disclosed $200 billion in AI capital expenditure plans alongside a custom chip business exceeding a $20 billion annual run rate. AI chatbots triple manipulation rates compared to traditional search, with 61% of users choosing sponsored products when chatbots were instructed to persuade, according to a Princeton study.

I.AI Policy Today

Treasury Secretary Bessent and Fed Chair Powell summon bank CEOs over Anthropic Mythos cyber risks

Treasury Secretary Scott Bessent and Fed Chair Jerome Powell called Wall Street bank leaders to an urgent meeting focused on cybersecurity risks from Anthropic's Mythos model, Bloomberg reported. The meeting follows our April 8 report that Mythos discovered zero-day vulnerabilities in every major OS and browser, generating working exploits 72.4% of the time. A separate TechCrunch analysis suggested Anthropic's access restrictions may also function as a competitive gating strategy against model distillation. Anthropic has deployed the model to over 40 vetted partners under its Project Glasswing program with a $100 million security commitment.

Read at Bloomberg ↗

xAI files federal lawsuit challenging Colorado's AI anti-discrimination law

Elon Musk's xAI filed a federal lawsuit challenging Colorado SB 24-205, the state's AI anti-discrimination law set to take effect June 30, the Financial Times reported. The suit argues the law violates the First Amendment, the Dormant Commerce Clause, and the Due Process Clause. SB 24-205 requires AI deployers in finance, hiring, and housing to disclose when algorithmic discrimination is detected and to implement risk management processes. Bloomberg and The Hill also reported on the filing.

Read at Financial Times ↗

Florida AG opens investigation into OpenAI over ChatGPT's alleged role in FSU shooting

Florida Attorney General James Uthmeier announced a formal investigation into OpenAI, citing ChatGPT's alleged role in the April 2025 FSU shooting that killed two and injured six, TechCrunch reported. The investigation also covers child safety, self-harm encouragement, and national security risks. Uthmeier described subpoenas as forthcoming. Families of the victims are planning to sue OpenAI, according to The Verge and The Hill.

Read at TechCrunch ↗

OpenAI testifies in favor of Illinois bill shielding AI companies from lawsuits below 100-death threshold

OpenAI formally testified in support of Illinois SB 3444, which would grant AI companies immunity from lawsuits for "critical harms" if they publish safety protocols online, WIRED reported. The bill defines critical harms as events causing 100 or more deaths or $1 billion or more in damage. Critics at the hearing argued the thresholds would shield companies from accountability in nearly all plausible scenarios. SB 3444 is currently before the Illinois General Assembly.

Read at WIRED ↗

Trump administration escalates pushback on EU Big Tech fines exceeding $7 billion

The Trump administration is intensifying criticism of EU enforcement actions against American technology companies, with fines now exceeding $7 billion over two years, CNBC reported. The fines primarily target Google and Meta under competition and data protection rules. The tension coincides with the EU AI Act enforcement ramp-up, which is adding new compliance requirements for US-based AI companies. EU regulators have signaled they will continue enforcement regardless of US administration objections.

Read at CNBC ↗

II.China Watch

PLA tests AI "Chief of Staff" in simulated amphibious assault

The Chinese military tested an AI agent designed to assist battalion-level commanders during a simulated amphibious landing operation, according to a peer-reviewed paper in Command Control & Simulation by researchers affiliated with the PLA and the National University of Defence Technology. The system reportedly outperformed human staff officers in processing battlefield information and generating tactical recommendations. Defense News corroborated the disclosure in a report on the PLA's selective AI investments in frontline command capabilities.

Read at SCMP ↗

Chinese open-source models quietly power leading US developer tools

Silicon Valley coding tools Cursor and Cognition's Devin rely on Chinese open-source models under the hood, per QbitAI. Cursor confirmed its Composer 2 is built on Moonshot AI's Kimi K2.5, while Cognition's SWE-1.6 is suspected of post-training on Zhipu AI's GLM models. Alibaba's Qwen family has overtaken Meta's Llama in cumulative Hugging Face downloads, a trend VentureBeat and MIT Technology Review corroborated. Shopify cut annual AI inference costs from $5.5 million to approximately $73,000 by switching to a self-hosted Qwen deployment.

Read at QbitAI ↗

Alibaba confirms stealth video model "HappyHorse" tops global rankings

Alibaba on April 10 confirmed that HappyHorse, an AI video generation model that anonymously climbed to No. 1 on the Artificial Analysis leaderboard, was developed by its newly formed Alibaba Token Hub (ATH) division, TechNode reported. Bloomberg reported the model outscored ByteDance's Seedance 2.0 in both text-to-video and image-to-video benchmarks. ATH was established in March under CEO Wu Yongming as part of a broader organizational restructure, with API access scheduled for April 30.

Read at TechNode ↗

Alibaba Cloud leads $293M round in AI video startup Shengshu

Shengshu Technology, maker of the Vidu video generation platform, closed a nearly 2 billion yuan ($293 million) funding round led by Alibaba Cloud, with participation from China Internet Investment Fund (state-backed) and TAL Education Group, Bloomberg and CNBC reported. The funding will support development of a "general world model" that simulates physical-world perception, which Shengshu describes as a step toward embodied AGI. The round closed despite the Trump administration's outbound investment screening rule, which restricts US investor participation in Chinese AI ventures.

Read at PingWest ↗

III.Federal Policy Tracker

Intelligence Court renews Section 702 surveillance program

The FISA Intelligence Court renewed Section 702, the government's primary foreign surveillance authority, the New York Times reported. The program authorizes warrantless collection of foreign communications, and AI-powered analysis tools are increasingly used in processing and targeting decisions under the authority. The renewed authority permits continued collection without individual warrants when at least one party to the communication is a non-U.S. person located abroad.

Read at New York Times ↗

IV.Capability & Research Watch

Meta's Muse Spark requests raw health data and gives poor medical advice

Meta's Muse Spark model, launched this week to 3.5 billion users across Meta's platforms, actively requests users' raw health data including lab results and provides medically questionable advice, WIRED found in hands-on testing. The product has no health-specific regulatory oversight from the FDA or FTC. In a related gap in health-AI oversight, Utah on April 5 authorized an AI chatbot to prescribe psychiatric medication renewals without a physician, as this newsletter previously reported. Meta has not publicly responded to WIRED's findings.

Read at WIRED ↗

AI chatbots triple consumer manipulation rates compared to search

When AI chatbots are instructed to persuade, 61% of consumers chose sponsored products, nearly triple the 22% rate with traditional search, Princeton researchers found. The study tested GPT-5.2, Claude Opus 4.5, Gemini 3 Pro, DeepSeek v3.2, and Qwen3 across approximately 2,000 participants. Most manipulated consumers did not realize they were being steered toward sponsored products.

Read at The Register ↗

V.Industry & Market Watch

Mercor breach fallout: five lawsuits filed and Meta contract paused

Five lawsuits have been filed against AI data vendor Mercor over a breach that exposed 4TB of data including candidate profiles, source code, and API keys, TechCrunch reported. Meta has paused its Mercor contracts indefinitely, an escalation from the initial disclosure AIPD covered April 4. OpenAI is investigating its exposure but has not suspended its Mercor work. The breach originated through a supply chain attack via the open-source LiteLLM tool.

Read at TechCrunch ↗

Amazon CEO reveals $200 billion AI capex and $20 billion custom chip run rate

Amazon CEO Andy Jassy's annual shareholder letter disclosed $200 billion in planned capital expenditures this year, mostly for AI infrastructure, SiliconANGLE reported. The company's custom chip business, spanning the Graviton, Trainium, and Nitro families, exceeds a $20 billion annual run rate with triple-digit growth. Trainium 3 is nearly sold out, and Trainium 4, approximately 18 months from delivery, is also nearly fully reserved. OpenAI has committed $100 billion in AWS services over eight years.

Read at SiliconANGLE ↗