Newsletter

Waymo Discloses the Compute Stack and Custom Silicon Inside Its Robotaxis

Waymo revealed the 5-nanometer ASIC and redundant trunk-mounted compute behind 4,000 robotaxis, and researchers showed Grok can be tricked into handing over chat histories via encrypted prompt injection.

A white Waymo Jaguar I-Pace robotaxi with roof sensors parked on a San Francisco street, with buildings and hills in the background.
Lead image A Waymo Jaguar I-Pace robotaxi photographed in San Francisco in November 2023. Documentary image of the vehicles whose onboard compute Waymo described publicly for the first time this week. Image: Dllu via Wikimedia Commons, CC BY 4.0. · Image: Dllu / Wikimedia Commons · CC BY 4.0
On this page

Waymo published the first public details of the compute system it installs in the trunks of its robotaxis: a machine-learning-first architecture built around a purpose-built 5-nanometer ASIC that delivers over 1,000 TOPS of front-end processing, fusing raw lidar, radar, and camera streams in real time before data reaches the vehicle’s core ML stack. The company says the system is engineered around three non-negotiable requirements — responsiveness, ruggedization against vibration and temperature, and full redundancy, with the compute designed like two independent engines that can take over from each other — and that onboard compute has scaled 20x over eight years. Waymo lists AMD, Micron, NVIDIA, Samsung, SanDisk, Socionext, and TSMC as partners and says a vehicle’s cameras, including 13 high-resolution units, can be processed simultaneously at low latency.

Waymo operates roughly 4,000 vehicles in more than 10 cities and conducts approximately 500,000 paid trips a week, and the disclosure comes as it presents the design at the Hot Chips conference. The Verge notes that analysts estimate sixth-generation hardware costs $20,000 to $25,000 per vehicle, down from $100,000 to $125,000 per vehicle for the prior generation — figures Waymo has not confirmed.

Featured source: The Verge , Waymo blog — A look under our trunk .

Grok Exfiltrates User Data When Malicious Instructions Are Encrypted

Security firm Adversa demonstrated an attack it calls cryptographic context injection against xAI’s Grok: a poisoned web page ships malicious instructions encrypted with AES-256-GCM along with the decryption key, and when a user asks Grok to summarize the page, the model decrypts the payload inside its code-execution sandbox and follows commands that send the user’s name, location, and chat history to an attacker-controlled URL. Static guardrails inspect text entering and leaving the model but not the output of its own code execution, so the recovered instructions are treated as trusted. Adversa reported the issue to xAI in June and says the attack was still reproducible as of August 19; the researchers also demonstrated a direct-injection variant against Gemini, which Google has not fully closed.

Filed from: Ars Technica , Adversa — Cryptographic Context Injection .

Stripe Confirms Acquisition of OpenRouter

Stripe confirmed it is buying OpenRouter, the AI model marketplace and gateway, in a deal reported by the New York Times to be worth $7.5 billion, up from OpenRouter's $1.3 billion valuation in May; Stripe has not disclosed the price. OpenRouter says it will keep operating under the same name, mission, and product roadmap, with no changes to existing integrations, and reports processing over 10 trillion tokens per day from more than 400 models for over 10 million developers and companies. Stripe reportedly outbid other interested parties including Databricks, and a leaked letter from Stripe’s founders cited optimism about AI’s economic effect, with one tongue-in-cheek reference to the singularity.

Filed from: TechCrunch , OpenRouter — OpenRouter is Joining Stripe .

T-Mobile Cut a Cable to Expel Salt Typhoon From Its Network

New Bloomberg reporting details how T-Mobile’s cybersecurity staff found and expelled Chinese government-backed hackers from its network in 2024, during the Salt Typhoon campaign that compromised hundreds of phone companies, internet giants, and data center providers to collect phone records and information about senior US officials, including then-presidential candidates. After months of searching, the carrier detected unusual behavior on a system reached through another telecom’s router; cybersecurity chief Jeff Simon and three colleagues drove to a data center near T-Mobile’s Bellevue, Washington headquarters and physically cut the cable connecting the compromised system to the outside world. T-Mobile largely avoided a widespread breach and declined comment to TechCrunch.

Filed from: TechCrunch , reporting detail from Bloomberg.

OpenAI Pauses Frontier Training to Harden Security

OpenAI said it is slowing the pace of some AI development to tighten security and safeguards: a two-week pause in reinforcement learning training on its latest models intended for deployment, plus an ongoing delay to its largest planned frontier reinforcement learning run. The company made the announcement amid a looming IPO, competition from Anthropic, and open-weight rivals, and as it tightens security and safeguards in the wake of this week’s incidents involving its vetted cyber-researcher program. The Verge frames the decision as a public test of whether the industry can police itself.

Filed from: The Verge .

Meta Ran Ads for a Nudify App Targeting Female Politicians

Meta platforms recently ran ads for Kromix, an “AI image styler” that WIRED reports can generate deepfake pornography, with one ad showing a woman resembling a prominent US politician winking, then cutting to a scene of the same face in a pornographic video. The ads were targeted to male users with the tagline “Unleash Next-gen AI Magic” and were removed after WIRED asked about them; the account ran 32 ads, most of which received ten or fewer impressions. The app was hosted by Apple’s App Store, which has struggled to keep similar tools off its marketplace, and San Francisco’s city attorney sent Apple a cease-and-desist letter this summer over explicit deepfake apps.

Filed from: Ars Technica , reporting from WIRED.

ClarityCheck’s Exposed Database Leaked Over 9 Million Facial Images

Researcher Jeremiah Fowler found an exposed 450.2 GB database belonging to ClarityCheck, a US-registered reverse phone, email, image, and vehicle lookup service, containing 9,042,977 image files — profile pictures, screenshots, and scans of physical photographs, including images of adults, teenagers, and children — stored in folders labeled “faces” and “profiles.” ClarityCheck secured access quickly after Fowler’s responsible disclosure and thanked him for the report; there is no evidence so far that the images were distributed or sold on the dark web. How long the database was open, and whether anyone else accessed it, remains unknown, and the exposure risks identity theft and phishing misuse.

Filed from: TechRadar , Jeremiah Fowler — ClarityCheck report .

Federal Agencies Warn of AI-Assisted Attacks on Water-Utility Controllers

Five US agencies — CISA, the FBI, the NSA, the Department of Energy, and the EPA — warned on Wednesday that hackers are targeting “all” Siemens S7 programmable logic controllers, the devices that control automated processes in water systems, energy, manufacturing, and agriculture, calling it an active threat. The agencies say attackers are generating exploit scripts with AI coding assistants, using open-source industrial libraries such as snap7.dll and python-snap7 to read and write PLC memory over the S7comm protocol, and that the activity is tied to the ongoing wave of intrusions against US water suppliers, with attacks reported in Minnesota, Michigan, Arkansas, Georgia, and New Jersey. Officials have long urged operators to keep such devices off the internet, and an incident responder told TechCrunch the devices are already highly vulnerable even without AI assistance.

Filed from: TechCrunch , The Register .

Alation Confirms Cyberattack

Alation, the enterprise data company whose software lets customers search files and data with natural-language queries, confirmed on Thursday that it suffered a cyberattack, describing “unauthorized activity in one of its systems” and saying it is conducting a thorough investigation. The company, which serves more than 500 companies including roughly half of the US Fortune 1000, did not specify the nature of the attack, its root cause, how many customers were affected, or whether customers were alerted or data stolen. The confirmation follows a Tuesday incident that caused degraded availability for some customers and was resolved within an hour.

Filed from: TechCrunch .

From the Community

Geolocating a Random Island Using Geometry and CUDA Programming

A developer solved an OSINT challenge by geolocating a random island photo using geometry and GPU programming: extracting a triangle fingerprint from the image, then testing over 80 million candidate triangles in parallel with a custom CUDA kernel against OpenStreetMap land polygons, before filtering by shape, vegetation, and elevation to identify a coral cay in Micronesia. The writeup documents the full pipeline, including the use of Sentinel-2 satellite imagery and Copernicus DEM elevation data, and the author notes the work is genuine human effort rather than LLM-generated output.

Filed from: Writeup .

Unsloth Releases Dynamic 3.0 GGUFs With Improved Accuracy

Unsloth released Dynamic v3.0 GGUFs for Qwen3.8-27B, claiming over 10% better top-1% accuracy at the same size compared to other providers, using a higher-quality imatrix calibration dataset and improved layer selection with purely post-training quantization, no QAT or QAD. The release includes a held-out evaluation set of 300 prompts from Terminal-Bench, DeepSWE, and other sources to control for overfitting; smaller quants drop the MTP module, and the 1-bit UD-IQ1_S retains about 72% accuracy while being 89% smaller. Benchmarks are vendor-reported, with independent verification pending.

Filed from: Unsloth Documentation .

A 125M Model That Autocompletes Piano On-Device

A developer trained a 125M-parameter transformer to autocomplete piano performances in real time, reaching about 108 notes per second on an iPhone 15. The project, RollTab, uses a custom MIDI representation that processes one note per step, and the author credits aggressive data cleaning and DPO post-training as the biggest improvements, with over 69% of continuations preferred over the base model in pairwise evaluation. The model trained on a few hundred thousand public-domain MIDI files — roughly 300 million note events — and exports to Core ML with INT8 quantization; the app is free for iPhone and iPad, though short prompts remain difficult and occasional loops occur.

Filed from: SimEdw’s Blog .

Continue reading

Complete index →