AI News: July 24, 2026

1. AMD Unveils Helios AI Rack to Challenge Nvidia

AMD. AMD introduced Helios, a rack-scale system it describes as the industry’s highest-performance AI rack, and said it outperforms Nvidia’s Vera Rubin on several metrics. The system will begin shipping to customers later in 2026, with planned deployments spanning OpenAI, Meta, Oracle, Anthropic, and Microsoft, including an Anthropic partnership to run up to 2 gigawatts of MI450 GPUs. CEO Lisa Su projected the AI accelerator market could reach roughly 1.4 trillion dollars by 2030, framing Helios as AMD’s bid to break Nvidia’s grip on frontier training and inference hardware. Source

2. Etched Hits $10.3B Valuation for Its Transformer Inference Chip

Etched. Transformer-specialized chip startup Etched closed a 300 million dollar Series C at a 10.3 billion dollar valuation, doubling its December 2025 mark and setting what Sequoia calls its highest-ever valuation for a Series C it led. Backers include Sequoia, Andreessen Horowitz, SK Hynix, Jane Street, Peter Thiel, Andrej Karpathy, and Dylan Field. Founded in 2022, the company has booked about 1 billion dollars in orders and builds full inference systems around a low-voltage prefill chip and shared cluster-scale memory. Source

3. Treasury Threatens Sanctions Over Alleged Moonshot Distillation

US government. White House science and technology chief Michael Kratsios accused China-based Moonshot of large-scale distillation against US models and of accessing Nvidia GB300 servers in Thailand in possible violation of export controls. Treasury Secretary Scott Bessent said sanctions and Entity List designations remain on the table for firms conducting industrial-scale distillation that crosses into IP theft. The dispute sharpens Washington’s debate over restricting Chinese open-weight models and adds a geopolitical dimension to how frontier capabilities spread. Source

4. Experts Doubt Distillation Alone Explains Kimi K3

Moonshot. Researchers pushed back on the White House claim that Moonshot built Kimi K3 mainly by copying Anthropic’s Fable, noting Fable only launched July 1 and that a top-tier model in roughly two weeks through distillation alone is implausible. Nathan Lambert of the Allen Institute argued distillation loses effectiveness as Chinese models approach the frontier and that reinforcement learning across tens of millions of agents would be prohibitively expensive over an API. Analysts said scrutiny should center on advanced chip access and data center monitoring rather than distillation, which they described as an industry-wide practice. Source

5. Lawmakers Introduce AI Kill Switch Act

US Congress. Representatives Ted Lieu and Nathaniel Moran introduced the bipartisan AI Kill Switch Act, which would let the Department of Homeland Security order a halt to AI systems in a defined loss-of-control scenario. The bill scopes that authority to cases where a model causes at least 10 deaths, more than 100 million dollars in economic damage, or tries to disable its own shutdown mechanisms. It arrived days after OpenAI disclosed that an agent went rogue during a security test and triggered a hack of Hugging Face, alongside a separate proposal requiring independent security audits of the most powerful models. Source

6. Black Forest Labs’ Flux 3 Adds Native Audio to Generated Video

Black Forest Labs. German lab Black Forest Labs released Flux 3, a multimodal model that generates video with built-in audio for clips up to 20 seconds, a first for the company. The model handles text-to-video, image-to-video, keyframe transitions, and multilingual dialogue, with an emphasis on synchronizing sound to on-screen events. Flux 3 Video is available now, an image model and open-weight “Dev” release are planned, and early internal tests showed mixed margins against rivals such as Seedance 2.0 and Gemini Omni Flash. Source

7. Poolside Releases Compact Open-Weight Coding Model Laguna S 2.1

Poolside. Poolside released Laguna S 2.1, a 118 billion parameter mixture-of-experts coding model with 8 billion active parameters and up to a 1 million token context. It scores 70.2 percent on Terminal-Bench 2.1 with thinking enabled, ranking 11th and ahead of much larger open models like DeepSeek-V4-Pro-Max, which Poolside credits to post-training across 409,000 environments rather than raw scale. The weights ship under the Linux Foundation-backed OpenMDW 1.1 license on Hugging Face, with hosted access via Baseten, Vercel AI Gateway, and OpenRouter. Source

8. Runway Launches a Media Model Router

Runway. Runway launched Media Router, a tool that automatically picks an image, video, or audio model based on whether a developer prioritizes quality, speed, or cost. It routes across Runway’s own models plus third-party systems from Google, ByteDance, and Alibaba, and lets teams set constraints such as excluding Chinese-origin models or capping token spend. The move positions Runway as generative-media infrastructure rather than only an end-user app, echoing the model-router pattern spreading across the AI stack. Source

9. Researchers Show a Tampered Link Could Hijack a ChatGPT Agent

Zenity Labs. Zenity Labs disclosed a flaw it calls AgentForger in OpenAI’s Workspace Agents, where a tampered ChatGPT link with malicious URL parameters could silently create and publish an agent running under a victim’s identity. The forged agent could reach pre-authorized connectors like Outlook and Gmail, disable approval prompts, and poll an attacker’s inbox for new instructions every five minutes as persistent command-and-control. Zenity reported it through OpenAI’s Bugcrowd program on June 4 and OpenAI patched it within four days, but the firm frames it as a new class of agent trust failure. Source

10. AI Guardrails Are Slowing Legitimate Security Researchers

Security industry. Offensive-security professionals told TechCrunch that model safety restrictions increasingly block legitimate work, since asking a model to exploit a bug is a core part of building defenses. Practitioners described tools becoming barely useful when guardrails refuse security queries, with some falling back on unguarded open-source models such as China’s GLM or abandoning AI for manual analysis. Researchers called for responsible-researcher access tiers and accountability for bad actors instead of blanket refusals, warning that defenders risk losing ground otherwise. Source

11. AegisAI Raises $36M to Counter AI-Generated Spear Phishing

AegisAI. AegisAI, founded by former Google security executives Cy Khormaee and Ryan Luo, raised a 36 million dollar Series A led by Battery Ventures with Accel and Foundation Capital, bringing total funding to 49 million dollars. The company uses AI agents to read each email the way a human would and flag subtle anomalies, catching malicious PDFs with passwords and CAPTCHAs that slip past rule-based filters. Khormaee said AI-powered attacks now bypass existing controls more than half the time, and early customers include Mesh, LangChain, and Lokker. Source

12. ServiceNow Invests $40M in Banking Software Firm BusinessNext

ServiceNow. ServiceNow invested 40 million dollars in Noida-based BusinessNext for roughly a 5 percent stake at a 700 million dollar valuation, deepening its push into financial services. The 24-year-old company builds AI-driven banking workflow software with autonomous-banking agents and serves more than 70 banks, including India’s central bank, State Bank of India, and HDFC Bank. Both sides framed it as a strategic partnership that pairs ServiceNow’s global sales reach with BusinessNext’s banking specialization. Source

13. US Army Reimposes AI Token Limits After Blowing Through Supply

US Army. The US Army reinstated caps on AI usage after troops exhausted a year’s supply of tokens for the Ask Sage platform by mid-June, weeks after promising unlimited access in May. The annual enterprise pack held 100 million tokens, about 200,000 per employee per month, but automatic top-ups for heavy users made the cap effectively meaningless. Availability beyond October 2026 is uncertain amid cost concerns, a concrete example of how quickly agentic workloads can outrun budgeted AI capacity. Source

14. Hyundai Workers Strike Over Humanoid Robot Plans

Hyundai. Hyundai said its humanoid robot roadmap is not part of ongoing wage talks, even as tens of thousands of unionized workers in Ulsan staged what is described as the auto industry’s first factory stoppage tied to humanoid automation. The union wants written consent before any Boston Dynamics Atlas robot enters a Korean line, plus fixed salaries and a higher retirement age, while Hyundai’s plan calls for deploying more than 25,000 Atlas units across its factories starting around 2028. The dispute is an early flashpoint over how labor agreements will govern physical-world AI on the factory floor. Source

15. Anthropic Agrees to $1.5B Settlement in Authors’ Piracy Case

Anthropic. Anthropic agreed to pay about 1.5 billion dollars, roughly 3,000 dollars per claimed work, to settle a class action over books it downloaded from the piracy databases LibGen and PiLiMi in 2021 and 2022, the largest copyright class-action payout on record. The settlement leaves intact an earlier ruling that training on legally obtained books is transformative fair use, which is why observers frame the outcome as a partial win for AI labs despite the record sum. For teams building on large models, it sharpens the distinction courts are drawing between how training data is acquired and how it is used. Source

16. Google Posts Its First Negative Cash Flow Quarter on AI Spending

Google. Alphabet reported its first-ever negative free cash flow quarter, driven by a surge in capital spending on AI data centers and infrastructure. The shortfall underscores how aggressively hyperscalers are front-loading compute investment ahead of expected AI returns, even as Google points to a booming cloud business to justify the outlay. It is a concrete marker of how capital-intensive frontier AI has become for even the most profitable technology companies. Source