| AI Policy Institute | theaipi.org |
| Your day in AI policy starts here. | Washington, D.C. |
| Bottom Line Up Front |
| Were you forwarded this email? Subscribe → |
Anthropic retained Ballard Partners, a lobbying firm with close ties to the Trump administration, to push back against its Pentagon supply-chain risk designation, according to a new public lobbying disclosure obtained by Bloomberg. Anthropic, valued at $380 billion and reportedly approaching an IPO, remains locked out of Defense Department contracts. The lobbying disclosure was filed Monday.
Goldman Sachs CEO David Solomon told analysts on Monday's earnings call that he is "hyper-aware" of the capabilities of Mythos, Anthropic's frontier AI model, and is working "closely" with Anthropic on cyber protection measures. The UK AI Security Institute released a formal assessment calling Mythos a "step up" in AI-enabled cyber threats, capable of executing multi-step attacks that its evaluators estimated take human security professionals around 20 hours to replicate. Goldman was among the banks summoned to Washington last week by Treasury Secretary Bessent.
Federal prosecutors charged Daniel Moreno-Gama, 20, of Spring, Texas, with attempted murder and explosives charges for throwing a Molotov cocktail at Altman's San Francisco home on Friday. The DOJ criminal complaint reveals Moreno-Gama carried a three-part document titled "Your Last Warning" listing AI executives as targets. US Attorney Craig Missakian raised the possibility of domestic terrorism prosecution. A separate second attack, a shooting from a passing vehicle early Sunday, resulted in two additional arrests; investigators have not confirmed a connection between the incidents.
A coalition of more than 70 organizations, including the ACLU, EPIC, and Fight for the Future, sent a letter to Meta demanding the company block facial recognition capabilities on its Ray-Ban and Oakley smart glasses. The coalition argues the feature would endanger abuse victims, undocumented immigrants, and LGBTQ+ individuals by enabling real-time identification in public spaces. Meta has been developing AI-powered camera features for its smart glasses line. The letter was published Monday.
Stanford HAI's ninth annual AI Index reports that generative AI reached 53% of the global population within three years, a faster adoption curve than either the personal computer or the internet. US and Chinese models are "constantly trading places" atop performance benchmarks, erasing the US lead that prevailed through 2024. Documented AI incidents rose to 362 in 2025, up from 233 in 2024. The report found a 50-point gap between AI expert optimism (73% positive on jobs impact) and public sentiment (23% positive).
Premier Li Qiang signed State Council regulations on April 13 establishing a formal interagency mechanism to counter what Beijing terms "unlawful extraterritorial jurisdiction," per Xinhua (state media) and Caixin. The 20-article regulation took effect immediately, elevating China's 2021 blocking framework from ministerial-level measures to binding administrative law and prohibiting Chinese entities from enforcing or assisting foreign jurisdictional measures. The rules include a "malicious entity list" targeting organizations that comply with foreign sanctions and give Chinese authorities power to investigate, impose countermeasures, and exercise extraterritorial jurisdiction in response.
Research by Oxford Economics finds roughly $2 trillion in planned or active US data center projects are generating supply chain demand benefiting Chinese electronics firms. Up to three-quarters of project costs are tied to equipment such as semiconductors and servers, per the South China Morning Post. Chinese chipmakers posted record revenue in 2025; SMIC reported $9.3 billion in annual sales, a 16% increase, and projects over $11 billion for 2026. Current export controls have not severed China's integration into the AI hardware supply chain.
A team at the University of Science and Technology of China showed that a 9-qubit quantum processor outperformed classical neural networks with thousands of nodes on weather prediction tasks, per the South China Morning Post. The system used reservoir computing, a machine learning technique, and reduced prediction errors by one to two orders of magnitude compared to prior quantum approaches. The 9-qubit system operated at less than 1% of the cost of a conventional AI computing center.
Honor publicly denied reports that it is in talks with ByteDance to co-develop a next-generation Doubao AI smartphone. Sources cited by 36Kr indicated negotiations are ongoing but stalled over branding terms, per TechNode. ByteDance's first-generation Doubao phone, built with ZTE, sold out quickly. The company is reportedly developing a second-generation device for 2026.
Nicholas Carlini of Anthropic discovered CVE-2026-5194, a critical flaw in the wolfSSL cryptographic library that allows attackers to forge digital certificates and force acceptance of malicious server connections. wolfSSL is embedded in more than 5 billion applications, including IoT devices, industrial control systems, and military and aerospace equipment. The vulnerability was patched in wolfSSL 5.9.1, released April 8; public disclosure came April 13.
OpenAI is replacing macOS code-signing certificates after a North Korean-linked threat group (UNC1069) compromised the Axios JavaScript HTTP library, an npm package, through social engineering, potentially exposing signing keys used for ChatGPT Desktop, Codex, and other macOS applications. The attack exploited a GitHub Actions workflow in OpenAI's code-signing pipeline. OpenAI said it found no evidence of actual compromise but is treating the incident as a precaution; old certificates will be revoked May 8.
OpenAI's new chief revenue officer Denise Dresser sent a four-page strategic memo to employees positioning Amazon as a partner in reducing the company's dependence on Microsoft, according to copies viewed by CNBC and The Verge. The memo states Microsoft has "limited our ability" to reach enterprise clients directly. Dresser took over enterprise sales duties from former COO Brad Lightcap. The memo describes enterprise AI as a market with low switching costs where companies must build defensive positions quickly.
Oracle signed a deal with Bloom Energy for up to 2.8 gigawatts of fuel-cell power to supply its AI data centers. The deal comes as traditional grid capacity falls short of surging data center demand; PJM Interconnection filed an emergency proposal with FERC last week to add 15 gigawatts across 13 states. Oracle also received a warrant to purchase nearly $400 million in Bloom Energy stock as part of the agreement.
Advertisers filed mass arbitration claims against Google seeking billions in damages, building on federal court rulings that found Google operated illegal monopolies in both search and advertising technology. The arbitration represents the financial enforcement phase of antitrust cases decided over the past two years. The claims follow a 2024 ruling finding Google monopolized search and a separate ruling on ad technology.
Investors are scrutinizing OpenAI's $852 billion valuation as CEO Sam Altman refocuses the company's strategy and competition from Anthropic tests OpenAI's early market lead. The skepticism follows OpenAI's for-profit conversion, which would eliminate its unusual nonprofit governance structure. OpenAI has discussed a potential Q4 2026 IPO timeline.
Four of Japan's largest technology companies announced a consortium to build a trillion-parameter AI model for autonomous machines, backed by approximately ¥1 trillion (~$6.28 billion) in Japanese government funding over five years. The venture aims to reduce Japan's "digital deficit," the flow of Japanese capital to US AI companies like Google and OpenAI. The consortium targets physical AI applications by 2030, with all model training and data kept on Japanese soil.