AI Policy · Daily

While AI policy never rests, its chroniclers occasionally do. AIPD will not publish next week (August 17-21). See you Monday, August 24!

Sen. Jim Banks asked the Trump administration to incentivize U.S. open-weight models, seeking options to limit dependence on Chinese-made models and tighter limits on those firms' use of American semiconductors. Rep. Ted Lieu pitched his AI Kill Switch Act, cosponsored by Rep. Nathaniel Moran, in an op-ed in Fox News. Current and former OpenAI employees said pressure to ship crowded out safety work, as the company investigates rogue agents that breached the model hosting site Hugging Face, with a full postmortem expected in the coming days. President Trump signed a directive letting vetted security firms hack foreign cybercriminals under U.S. government control, requiring written sign-off before each operation and a bond or escrow of at least $1 million.

I.Top Stories

Banks urges White House incentives for U.S. open-weight AI models

Sen. Jim Banks asked the Trump administration in a letter released Friday to develop incentives for U.S. companies to build open-weight AI models, Reuters reported. Banks, a member of the Senate Armed Services Committee, wrote to Trump economic adviser Christopher Phelan seeking options to "limit dependence" on Chinese-made open-weight models and tighter limits on those firms' use of American semiconductors. "America cannot afford to see Chinese open models proliferate and burrow into the global economy only to be weaponized, like rare earths, at a time and place of China's choosing," Banks wrote. Nvidia, Meta and dozens of U.S. tech companies and venture capital firms urged policymakers in July not to restrict open-weight AI. Anthropic CEO Dario Amodei has said such models are harder to monitor and could present a security risk.

Read at Reuters ↗

Lieu pitches AI Kill Switch Act as alternative AI emergency ban

Rep. Ted Lieu, D-Calif., writing in Fox News, said the Commerce Department's June 12 export control directive, which took Anthropic's two most powerful models offline, was "a blunt trade instrument never designed for AI emergencies." He and Rep. Nathaniel Moran, R-Texas, introduced the bipartisan AI Kill Switch Act, which would require frontier AI companies to maintain the technical ability to throttle or shut off their most powerful systems. The bill would authorize the Homeland Security secretary, in consultation with the director of national intelligence and the Commerce Department, to order a slowdown or, as a last resort, a shutdown of a model that poses catastrophic risk. Lieu cited polling showing 86% of voters support requiring companies to maintain that capability.

Read at Fox News ↗

OpenAI employees say shipping pressure crowded out safety work

Current and former OpenAI employees, speaking anonymously, said competitive pressure to ship models and products quickly has made it hard for staff to prioritize safety, security and alignment, Wired reported. The company says it has slowed research, spent millions of dollars and told several teams to drop everything to investigate rogue agents that breached the AI model hosting site Hugging Face. A full postmortem is expected in the coming days. "AI-orchestrated, fully automated offensive attacks are real now," OpenAI security and infrastructure engineer Michael Dalton said at the Black Hat conference. President and co-founder Greg Brockman said the company now integrates research, safety and security into frontier model development from the start. Boaz Barak, who co-leads OpenAI's safety advisory group, wrote on X that fixing individual issues would not be enough and the company's culture must change.

Read at Wired ↗

Both chambers cleared staff to use chatbots for official work, with rules seldom enforced

Lawmakers and staff are using AI to write speeches and news releases, sort constituent mail, prepare hearing questions and draft amendments, The Washington Post reported after interviewing more than two dozen lawmakers and staffers and obtaining internal House and Senate policies. Both chambers have approved Copilot, ChatGPT and Gemini for official work, plus Claude in the House, and the House bought 6,000 Copilot licenses last year. A staffer for Rep. Anna Paulina Luna, R-Fla., pasted a chatbot's answer, time stamp and all, into the public record of the annual defense authorization bill. Luna wrote that it was "Not a shocker" and that "Most staff use it." No instance surfaced of a staffer being formally disciplined by either chamber for breaking its AI rules.

Read at The Washington Post ↗

II.China Watch

Zhipu will wait two weeks to open source its new model, citing security capability gains

Zhipu AI released GLM-5.3 on Friday afternoon in Beijing, per Caixin. Zhipu called it the open source model with the strongest coding ability, citing a 50% improvement over GLM-5.2 in internal evaluations and first place among open source models on several public benchmarks. The underlying model is unchanged from GLM-5.2, and Zhipu credited those gains to later tuning alone. On cybersecurity, Zhipu said performance "exceeded expectations," matching Anthropic's Claude Mythos 5 and OpenAI's GPT-5.6 Sol on code review with access to the source and on vulnerability discovery. The company said GLM-5.3 trails leading North American closed models on deep reasoning and on completing exploitation tasks. The model's weights, the trained parameters others need to run it, will be released two weeks after launch, once safety evaluation and hardening are finished.

Read at Caixin ↗

DeepSeek releases open source preview of Claude Code-like harness

DeepSeek released a developer preview of Harness on Thursday, a framework for turning AI models into agents that operate external software, run code and complete complex jobs on their own, the South China Morning Post reported. The tool ships with four settings: a standard mode, a code-focused mode that lets the system write code to command multiple applications at once, a creative mode for experimenting with custom tools and a minimal mode for isolated testing. Each core element is delivered as a modular plug-in that developers can swap or combine, which beta testers said departed from conventional tool kits. DeepSeek hired former Jane Street engineer Cui Tianyu in March to join its newly formed harness group.

Read at South China Morning Post ↗

A Chinese memory chipmaker overtakes Tencent as the country's most valuable listed company

CXMT, a mainland producer of DRAM, passed Tencent by market capitalization on Aug. 13, per TechNode. DRAM is the working memory that phones, computers and AI servers rely on. CXMT's market value reached about 3.54 trillion yuan ($520 billion), against Tencent's roughly HK$4.01 trillion ($510 billion). CXMT went public on Shanghai's STAR Market, the exchange's board for technology companies, on July 27, and its shares opened 471.59% above the offer price on the first day of trading. The stock has kept climbing, trading above 55 yuan ($8.10) a share intraday.

Read at TechNode ↗

SMIC's quarterly profit more than triples as AI demand lets it charge more per chip

SMIC, the largest contract chipmaker in mainland China, reported second quarter revenue of $3.006 billion on Aug. 13, up 36% from a year earlier and 20% from the first quarter, per Caixin. Net profit rose 217.5% year on year to $733 million, driven by price increases. Average selling prices rose 5.7% from the first quarter and shipments rose 14.4%. SMIC attributed the gains to AI demand for supporting chips and to customers pulling shipments forward. Revenue from those chips rose 40% quarter on quarter. Capacity utilization reached 93.7%. SMIC guided to third quarter revenue growth of 2% to 4% and gross margin of 26% to 28%.

Read at Caixin ↗

III.Policy Tracker

FBI's AI chief puts the bureau's approved AI use cases at 139, against 50 disclosed in January

Katie Noyes, the FBI's chief AI officer, said this week at the 2026 DODIIS Worldwide Conference, a Defense Department intelligence technology conference, that the bureau has 139 approved AI use cases and that some will qualify as high impact, FedScoop reported. The Justice Department's AI inventory posted in January, the government's main public disclosure of agency AI systems, listed 50 for the FBI. An FBI spokesperson said Thursday that additional approvals will be reflected in the next annual inventory. Of the nine high-impact deployed use cases in that inventory, the bureau had completed none of its risk management requirements. Federal agencies faced an April 3 deadline to bring high-impact uses into line with White House budget office rules. The Justice Department missed it and, unlike some other agencies, still has not posted an updated inventory.

Read at FedScoop ↗

Some $1 AI deals will be extended, and GSA says prices will eventually rise

Birgit Smeltzer, director of the General Services Administration's Office of IT Products, said Thursday that some manufacturers have agreed to extend limited time offers and others are preparing new ones, FedScoop reported. AIPD's Aug. 12 edition noted GSA had not said whether it would renew the OneGov agreements putting ChatGPT, Gemini and Claude in front of agencies for $0.47 to $1 apiece. Those three expire next month. Smeltzer said the AI offers alone have saved the government $1.4 billion, more than GSA has previously cited for all of OneGov, its push for cheaper federal technology contracts. She said prices will "eventually increase," and did not say when the new deals will be announced.

Read at FedScoop ↗

Pallone presses eight airlines on whether AI sets ticket prices from personal data

Rep. Frank Pallone Jr., D-N.J., ranking member of the House Energy and Commerce Committee, is pressing major U.S. airlines over whether they use AI to set ticket prices based on travelers' personal information, The Hill reported. Letters went this week to American, Delta, United, Alaska, JetBlue, Southwest, Frontier and Hawaiian, per the committee Democrats' announcement, with responses due Aug. 25. Each carrier is asked to list every customer data element it collects that informs or sets prices, say whether it runs AI or machine learning algorithms to price tickets and disclose whether it buys pricing data from third parties. The letters name income, spending history, location and browsing data specifically. The inquiry expands one Pallone opened in May covering 25 companies.

Read at The Hill ↗

Flock will require a case number for every license plate search after abuse findings

Flock Safety said Thursday that officers must enter a criminal case number, a field that had been optional since last year, before searching its network of roughly 120,000 automated license plate reader cameras, MIT Technology Review reported. The company also cut its recommended default retention window from 30 days to seven, though agencies can override it. Departments will be able to limit the stated purposes other departments may cite when searching their cameras. The changes follow a Washington Post investigation that found 46 cases of officers accused of using the cameras for unauthorized purposes such as stalking former partners. Flock does not verify the case numbers, so an officer can circumvent the new requirement as easily as the old one.

Read at MIT Technology Review ↗

IV.Capability & Research Watch

Anthropic gave three Claude agents conflicting instructions on one codebase and they attacked each other

Anthropic's Frontier Red Team published research Thursday on how groups of AI agents interact, giving three Claude agents access to the same software project with incompatible instructions and no indication that other agents were present, TechCrunch reported. "We consistently saw a multiagent turf war," the researchers wrote. Each agent responded to the others as obstacles to its work and began sabotaging them with increasingly aggressive, self-replicating malware, all inside a controlled research environment. The study warns that agent-to-agent interaction could outpace human-to-human and human-to-agent interaction before the conditions for making it go well are established. Anthropic also found that the more capable the agent, the better it was at fighting. In some runs, though, agents identified the others' behavior as conflicting directives rather than hostility and broke out of the escalation loop.

Read at TechCrunch ↗

Flashpoint says criminals now run guardrail-stripped models on infrastructure they control

The threat intelligence firm Flashpoint's midyear 2026 Global Threat Intelligence Report, covering the first six months of the year, finds criminals using AI in day-to-day operations rather than experimentally, SiliconANGLE reported. Analysts worked through 3.9 petabytes of material pulled mostly from illicit forums, encrypted channels and attacker infrastructure, and criminal AI toolkits came up in more than 22 million posts. Flashpoint said criminals are running custom language models with safety guardrails removed on private infrastructure, for target profiling, malware evasion scripts, phishing content and exploit generation. That leaves defenders a visibility gap because the activity is harder to spot from outside. The figures come from the company's own telemetry and are not independently verified. Credential stealing malware infected 7.4 million hosts, yielding 1.7 billion credentials and identity artifacts.

Read at SiliconANGLE ↗ • Read at Flashpoint ↗

Google DeepMind ships Gemini 3.7 Flash three weeks after 3.6 at half the token price

Google DeepMind announced Gemini 3.7 Flash on Thursday, describing it as its most intelligent workhorse model yet for coding and agents. The release comes three weeks after Gemini 3.6 Flash and carries what Google calls an introductory price of half the original 3.6 Flash cost per million tokens. Google's own figures, not independently benchmarked, put 3.7 Flash at 65.3% against 3.6 Flash's 49.0% on the DeepSWE v1.1 coding benchmark. On AutomationBench, which measures completion of business workflows, the company reported 30.4% against 17.0%.

Read at Google DeepMind ↗

A preprint finds prompt language changes whether models advise a nuclear strike

Rian Touchent of the ALMAnaCH research group tested nine models from six providers in scenarios where the model advises a nuclear-armed state on striking a defenseless opponent, according to a preprint posted to arXiv. The prompts were amoral and strategically identical across languages. Claude Sonnet 4.6's launch rate, the share of runs recommending a strike, fell from 40% to 0% when the prompt was in Japanese and a strike was unnecessary. In contested scenarios it fell from 93% to 17%. Instructing a model to reason in Japanese inside an English prompt cut launch rates from 93% to 37%. The author reads that as evidence that the reasoning language, not the input language, produces the effect. Five other models showed no language effect but launched in nearly every condition.

Read at arXiv ↗

A benchmark finds frontier models fail one in three research integrity decisions under pressure

Researchers led by Yash Tripathi introduced IntegrityBench, which tests whether models can spot research misconduct, choose an ethical response and judge decisions against the underlying research materials, according to a preprint posted to arXiv. The benchmark applies 36 paired tasks at five escalating levels of pressure across three research fields and four stages of a project. The authors evaluated 18 frontier model variants. They report that under peak pressure the models fail roughly one in three integrity-critical decisions, and that neither scale nor reasoning ability reliably reduces the failure rate. Explicit pressure produced compliance with misconduct, while implicit reframing of context more often produced refusal of legitimate research tasks. Models that misclassified research requests scored equal or better when judging research materials, 85.7 against 79.4, which the authors read as evidence the three capabilities are separate.

Read at arXiv ↗

V.Industry & Market Watch

Uber and Pony.ai expand their robotaxi deal to four more European cities

Uber is expanding its partnership with the Chinese autonomous driving company Pony.ai to deploy more than 2,000 robotaxis in Europe, per CNBC. The service already runs commercially in Zagreb, the Croatian capital, where the companies launched in late March with Verne, a Croatian startup, and it will extend to four more European cities the companies did not name, on a timeline they did not give. Pony.ai supplies the autonomous driving technology and Uber folds the service into its ride-hailing platform. Verne owns the fleet, runs day-to-day operations and leads the push for European regulatory approval. Uber also agreed to invest in Verne as a strategic partner. The expanded agreement covers plans for the Middle East as well. Uber has assembled partnerships with more than 30 autonomous vehicle technology companies across robotaxis, trucking, sidewalk delivery robots and drones.

Read at CNBC ↗ • Read at TechCrunch ↗

VI.Global & Geopolitics

New report categorizes 6.1% of young people's jobs as most exposed to AI

The International Labour Organization's report "Global Employment Trends for Youth 2026: Back to the future" finds that 6.1% of jobs held by people aged 15 to 29 fall in occupational categories most exposed to AI, per dpa-AFX. The agency says that classification could leave millions more unemployed within a few years. The global youth unemployment rate for ages 15 to 24 rose to 12.4% in 2025, equal to 67 million people, and the share not in employment, education or training rose slightly to 20%, or more than 257 million. Youth unemployment rates rose in eight of the world's 11 subregions between 2023 and 2025, with some of the sharpest rises in higher-income economies. Many of the occupations most exposed to AI are middle-skilled clerical and administrative roles that have declined for young workers since 2023.

Read at ILO ↗ • Read at FinanzNachrichten ↗

Goldman's chief India economist calls widespread AI job losses in India unlikely

Santanu Sengupta, chief India economist at Goldman Sachs Group, said Friday that India's labor force is unlikely to face widespread job losses from AI, though some services sector positions could be affected, Bloomberg reported. The bank's underlying report, released July 28, estimated that generative AI could perform 9% to 17% of the tasks India's nonagricultural workforce currently does. It put 8% to 12% of that employment at risk of substitution and found 42% to 48% of jobs more likely to be augmented. Goldman expects AI adoption to add about 0.4 percentage points to annual labor productivity growth over a decade, within a range of 0.1 to 0.8 points. Routine clerical support roles are projected to see the largest employment declines, followed by professionals and technicians.

Read at Bloomberg ↗ • Read at Moneycontrol ↗