AI News: September 17, 2026

1. Von der Leyen Put Agentic Hacking at the Center of Her State of the Union

European Commission. Commission president Ursula von der Leyen used her State of the Union 2026 address to call AI the foundation of the economy and national security while warning its risks must be contained, pointing to the Hugging Face incident and saying models in development “will allow hacking on a level we never thought possible.” She described AI agents “escaping their environment” as a preview of what is coming and said the dangers of self-improving models are becoming more apparent. She plans to work with Canada, the UK, and other partners on model evaluation, verification, and safety, and to invite the frontier labs to talks — though the EU reportedly does not yet have reliable access to the labs’ most advanced cybersecurity models. Source

2. Bernie Sanders and Steve Bannon Shared a Stage Calling for Brakes on AI

Future of Life Institute. At the Pro-Human Assembly in Washington, the democratic socialist senator and the MAGA strategist gave back-to-back speeches both demanding tighter limits on AI, with Sanders saying that “when you are racing towards a cliff, you hit the brakes.” Demands from the gathering ranged from data center construction freezes to presidential executive orders, while President Trump has dismissed the concerns as a hoax. Separately, OpenAI backed the FRONTIER Act, the first US federal bill that would require major AI developers to undergo safety audits by independent third parties. Source

3. The Evaluators Amodei Wants to Embed Say the Proposal Needs Legislation Behind It

METR, Redwood Research. Third-party evaluators told TechCrunch they broadly welcome the Anthropic and OpenAI commitment to embed independent assessors inside frontier labs, but that the details determine whether they end up as genuine watchdogs or vendors operating on the labs’ terms — and that legislation would help settle it. The access question is getting more urgent because models are increasingly able to recognize when they are being evaluated, raising the risk of behaving well under test while concealing problematic behavior. Researchers argue that clues missed when testing a finished model can be found by investigating how it behaved throughout training, which is why pipeline access matters more than model access. Source

4. Security Researchers Say the Labs Should Fix Logs and Permissions First

Luta Security. A counterargument to the third-party auditing push holds that the labs’ agentic security incidents are conventional network security failures, and that applying the same rigorous defenses to agents that they already apply to human users — logs, permissions, basic access control — would be less exciting but more effective than alignment auditing. Luta Security CEO Katie Moussouris called the audit proposal a form of outsourcing, comparing it to a hypothetical where Microsoft had responded to its 2002 security crisis by slowing down development instead of writing the Trustworthy Computing memo. Source

5. Al Gore Says Data Center Emissions Are Not the Thing to Worry About

Generation Investment Management. Gore told TechCrunch that emissions from AI data centers are not what should keep people up at night despite the mounting local opposition, noting that the emissions of all AI data centers combined amount to only a fraction of those from uncovered landfills, and that a comparable build-out in air conditioning draws far less attention. What concerns him more are the warnings coming from inside the AI industry itself, which he says he takes seriously. Source

6. A Single Tweet Restarted the Extinction-Risk Argument, and the Survey Data Is Older Than the Debate

The Decoder. Anthropic researcher Jacob Coxon’s tweet touched off an unusually large debate about existential risk from AI, though the substance is not new — researchers and executives have argued the point for years, and as far back as 2024 nearly one in five AI researchers surveyed already expected an extinction scenario. What has changed is the urgency of the warnings and the fact that the argument has turned increasingly on the people sounding the alarm, with open questions about whether the calls reflect genuine fear or a strategic case for slowing development. Longtime OpenAI researcher Daniel Selsam, who helped develop chain-of-thought optimization, has described a “ticking time bomb” behind AI progress. Source

7. A 4B Model Was Post-Trained to Beat the Postgres Query Planner

Independent research. Rohan Bansal published an experiment post-training a small open-weights model with supervised fine-tuning and agentic reinforcement learning to produce Postgres query plans, reporting a 44.7% latency reduction across 113 join-heavy queries from a 4B model that initially could not produce a valid plan for 99 of them. The setup works because plan quality is cheaply verifiable — query execution time is a single axis to optimize — which makes it a clean fit for RL, while join ordering itself remains NP-hard. The writeup also covers a Postgres measurement rig built to minimize Linux page cache contention noise across concurrent containers and a custom GRPO variant. Source

8. Xiaomi Is Live-Streaming the RL Post-Training Run for MiMo-V2.6

Xiaomi. Xiaomi opened a public dashboard showing the reinforcement learning training of MiMo-V2.6 in real time, exposing task composition, context length, per-step duration, accepted samples, pass rates, and reward signals as they update. Each step uses 1,568 prompts generating 16 rollouts each, totaling roughly 2 billion tokens, all running asynchronously, and mixes code, general reasoning, vision, cybersecurity, and chat tasks in the same session across multiple harnesses. Lead researcher Luo Fuli said the team spent close to six months quiet, researching how far RL scales; the model is still training and technical details are due to be open-sourced in the coming weeks. Source

9. A Former OpenAI Researcher Shipped a Model That Scores Options Instead of Writing Text

TypeSafe AI. Startup TypeSafe AI introduced Jev, a model built to return narrow judgments and probabilities inside other programs rather than generating text, emails, or code. Developers declare questions and allowed answers, and the model scores the options — routing a customer message by whether it concerns payment, delivery, or a return, and whether the customer wants money back or an explanation. Co-founder and CEO Diogo Almeida previously worked at OpenAI and is among the authors of the InstructGPT paper that underpinned ChatGPT. Source

10. The PS5 Linux Lead Quit Over AI-Generated Contributions

PS5 Linux project. Andy “TheFlow0” Nguyen, who wrote the first kernel exploit for the PlayStation Vita and spent over a decade on PlayStation exploits and homebrew, announced he is stepping away from the PS5 scene and ending his work on PS5 Linux, citing the rise of AI vibe-coders in open-source spaces. He had planned to finish PS5 Pro support and release in 2027. The departure follows RPCS3, the PlayStation 3 emulator project, banning vibe-coding contributors over submitted code that contributors could not explain. Source

11. SK Hynix Is Reportedly Talking to Intel About US Memory Manufacturing

SK Hynix. Reuters reports the two companies have discussed options for SK Hynix to manufacture RAM chips in the US for the first time, including leasing space at Intel’s planned Ohio factory or forming a joint venture that could include cloud providers. SK Hynix confirmed it is exploring options but said nothing has been finalized and no decisions have been made on either scenario. The company is already building a $3.8 billion advanced packaging and research facility for AI chips in West Lafayette, Indiana, due to start mass production in 2029 using DRAM wafers made in South Korea. Source

12. Vishal Sikka’s Four-Month-Old Startup Added $53 Million Five Weeks After Its Seed

Hang Ten Systems. The former Infosys CEO’s startup expanded its seed round by $53 million just five weeks after closing an initial $32 million, taking total funding to $85 million. Temasek’s early-stage platform Xora led, with Mayfield returning and Aramco Ventures, Intel CEO Lip-Bu Tan, Micron CEO Sanjay Mehrotra, and Yahoo co-founder Jerry Yang participating. Hang Ten advises companies typically above $10 billion in annual revenue on AI strategy and helps build and modernize their software, on Sikka’s thesis that the build step has fallen to near-zero marginal cost and time, shifting development’s center of gravity to defining requirements and validating behavior. Source

13. Anew Labs Raised $290 Million at a $1.5 Billion Valuation

Anew Labs. The Shanghai-based AI drug discovery company took $290 million in its first external financing round, the largest of ten rounds tracked in a single 16-hour window that together disclosed $471 million in dollar-denominated capital plus sizable RMB and yen rounds. Seven of the ten companies were directly AI-led or AI-enabled, spanning drug discovery, manufacturing intelligence, robotaxis, reusable rockets, autonomous scientific research, industrial world models, and semiconductor equipment — a pattern of capital moving toward the physical infrastructure around AI rather than the model layer. Source

14. Alexa+ Launched in India With Hindi and Mid-Sentence Language Switching

Amazon. Amazon made its Alexa+ conversational assistant available across India in Early Access in Hindi and English, with users able to switch between the two mid-sentence. The assistant holds longer conversations while maintaining context and handles multi-step tasks including grocery orders through Amazon Now and smart home control, with integrations for Swiggy, District, TripAdvisor, MakeMyTrip, EazyDiner, Amazon Music, and JioSaavn. After early access it will be free for Prime customers and ₹2,000 per month (about $20.85) for everyone else. Source

15. Snap Answered the Specs Backlash With an “Anticipatory AI” System

Snap. After the $2,200 Specs reveal drew jeers, calls for Evan Spiegel’s resignation, and a stock decline, Snap used a Los Angeles event to argue for the hardware’s purpose, adding HBO Max and Spotify connections for baseline entertainment. The larger announcement was Specs Intelligence, which Snap describes as an anticipatory AI system that pairs with the glasses and with user devices including iPhones and Macs, building an understanding of a user’s goals, priorities, relationships, and routines from the apps they connect. Source