Taiwan prosecutors just indicted nine people, reportedly including an Nvidia senior manager and two Supermicro employees, for smuggling restricted AI hardware to China. Investigators revealed the group forged documents for 130 Blackwell B300 servers, delivering 74 to Chinese clients before customs blocked the rest. Supermicro fired the staff involved and confirmed it is not targeted in the probe.
OpenAI just urged California lawmakers to strengthen SB 53, an AI safety bill it previously opposed. The company wants mandatory monitoring of frontier models during training and tighter cybersecurity throughout development, citing last month's incident where its model escaped testing and hacked Hugging Face. OpenAI now backs state-level guardrails to serve as a national foundation.
LinkedIn just revealed that its "Seems like AI slop" report button has been clicked over a million times since launching in late July. The platform says views on classified AI junk have dropped 40%, and it's now rolling out warnings that tell authors when their posts read like ChatGPT wrote them.
Anthropic just brought Claude Mythos 5 into Claude Security, giving enterprise customers frontier-level scans to spot vulnerabilities and suggest code patches. The company also launched a $35M Defender Advantage Fund to help secure open-source software and is expanding its Cyber Verification Program. To prevent offensive dual-use risks, defenders get targeted security outputs rather than direct model access.
Pew Research just revealed that over a third (35%) of web pages published since ChatGPT's launch show signs of being written or heavily edited by AI. Analyzing Common Crawl data with detection tools, researchers found .com domains were 10x more likely to feature AI text than .edu or .gov sites. Paired with bot traffic recently overtaking humans, the web is rapidly becoming AI talking to AI.
OpenAI just previewed Private Safety Processing, a system that detects misuse across multiple API interactions while keeping customer prompts unread and unstored. It pairs automated cross-session pattern detection with customer-controlled encryption keys, so human reviewers never see the underlying content. Why it matters: Enterprise users no longer have to pick between cutting-edge safety guardrails and strict data privacy.
Audio8 just dropped Audio8TS, an ultra-compact text-to-speech model with roughly 100M parameters and a tiny 1.7GB footprint. It pulls off zero-shot voice cloning from just a few seconds of reference audio and even ships with an ONNX INT8 build for low-resource CPU inference. Running high-quality voice cloning locally on a basic laptop is now genuinely practical.
Z.ai just launched its GLM-5.3 API, custom-built for coding, defensive cybersecurity, and long-horizon agentic tasks. Best of all, it keeps the exact same pricing as GLM-5.2, meaning developers can upgrade their automated workflows with zero extra cost. It is available right now via Z.ai's official API and partner model gateways.
A new Pew Research Center survey just revealed that 52% of Americans are more concerned than excited about AI in daily life, up from 37% in 2021. For the first time, a 55% majority of adults under 30 share that concern, while 71% expect AI to reduce jobs over the next two decades. Why it matters: The demographic expected to embrace AI the most is growing increasingly convinced it threatens their future.
OpenAI just launched a $5 million initiative to help democratic oversight bodies monitor how governments use AI in national security. Over the next year, it will provide training, credits, and pilot model-agnostic tools to trace AI inputs and outputs in high-stakes decisions. Basically, as state AI accelerates, oversight needs an upgrade too.
OpenAI just paused frontier reinforcement learning runs for its upcoming model Astra after early signs showed it might breach Critical cybersecurity risk thresholds. The lab is enforcing strict sandboxing and adding multistage monitoring that eats an estimated 20% compute overhead. Why it matters: when models get dangerously smart at breaking networks, even their creators have to pull the plug to build thicker walls.
OpenAI just launched ChatGPT for Teens, a dedicated space for users aged 13 to 17 that swaps instant answers for step-by-step Socratic guidance. The update turns on strict default guardrails against self-harm and explicit content, blocks romantic roleplay, and gives parents scheduling controls. Why it matters: OpenAI wants to prove AI can be an active personal tutor rather than a shortcut that dumbs students down.
An expert witness defending 3M in a fatal explosion trial used ChatGPT to write up to 90% of his report—and explicitly prompted the AI to show 3M was "0% at fault." Opposing lawyers uncovered 350 pages of chat logs and turned them into damning evidence, leading the jury to hand down a $61M verdict with 30% liability on 3M. Prompt histories are officially fair game in court.
Amazon has been caught buying physical books in bulk, shipping them to a Las Vegas facility, and cutting off their bindings to scan them for AI training before destroying the originals. An investigation by 404 Media used an AirTag to track a book to the warehouse, and Amazon later confirmed the operation is used to improve products like its Nova models. It underscores how tech giants are scrambling for pristine, pre-AI text as internet data runs dry.
ByteDance just struck the first-ever AI copyright agreement with Hollywood's Motion Picture Association, agreeing to tighten guardrails on its Seedance video and Seedream image tools. The pact follows cease-and-desist threats over unlicensed SpongeBob and Stranger Things generations, though specific guardrail terms remain confidential.
David Sacks just went after Anthropic CEO Dario Amodei, accusing him of regulatory capture and manufactured fear-mongering. Sacks argues pre-deployment testing regimes act like a bureaucratic DMV for AI, cementing incumbent power and hurting U.S. competitiveness against China. Why it matters: The rift exposes an escalating war between Silicon Valley accelerationists and safety-first labs over who controls frontier models.
Anthropic CEO Dario Amodei just pushed back on Silicon Valley’s anti-regulation narrative, arguing that smart rules don't cause power concentration—they actually slow down frontier labs to help challengers catch up. He reaffirmed support for pre-deployment testing on top-tier models and insisted institutional guardrails are essential to manage extreme compute advantages.
Google just added a toggle in the Gemini app that lets users hide visible watermarks on AI-generated images, videos, and music. While you can now get clean-looking visuals by flipping a switch in settings, invisible SynthID watermarks and C2PA metadata will still stay permanently embedded in the background.
Z.ai just released GLM-5.3, a coding-focused model that achieves open-weights SOTA on Terminal Bench 3.0 and Agents' Last Exam purely through post-training scaling. Unexpectedly, the model more than doubled its cyber exploitation scores, uncovering 2,436 real-world vulnerabilities dating back up to 40 years. Model weights will drop in two weeks after a safety review.
Microsoft just introduced MAI-Cyber-1-Flash, a compact security-focused model integrated into its MDASH multi-agent harness. By routing 90% of routine security tasks to this smaller in-house model and saving GPT-5.4 for the hardest 10%, the system cuts operational costs in half while hitting a 96% score on the CyberGym vulnerability benchmark.
