Samwise Tech/AI/Robotics Newsletter
Thursday, April 9, 2026
Anthropic Restricts New Claude Mythos Model to 40 Partners in Cybersecurity Push
Anthropic on April 9, 2026 unveiled Project Glasswing, a cybersecurity initiative built around Claude Mythos Preview—a powerful new AI model designed specifically to find software vulnerabilities. The model demonstrated capabilities far beyond existing tools, uncovering a 27-year-old vulnerability in OpenBSD and a 16-year-old flaw in FFmpeg that automated testing had missed for decades. Anthropic is restricting access to 40 vetted organizations including AWS, Apple, Cisco, Google, and Microsoft, citing the model’s dual-use risk. The company committed $100 million in usage credits and $4 million in donations to open-source security groups supporting the initiative.
Sources: TechCrunch
Meta Debuts Muse Spark, First Model from Superintelligence Labs, in Major AI Overhaul
Meta unveiled Muse Spark on April 8, 2026, its most powerful AI model yet and the first output of Meta Superintelligence Labs, led by Alexandr Wang, whom Meta recruited from Scale AI. Unlike Meta’s typical open-source releases, Muse Spark is proprietary. The model ranks fifth globally on the Artificial Analysis Intelligence Index with a score of 52, supporting multimodal input including voice, text, and images, plus health assistance, complex reasoning, and agentic capabilities. The model is available free at meta.ai and will roll out to WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban AI glasses over the coming weeks.
Sources: TechCrunch
OpenAI Opens Applications for Paid External AI Safety Research Fellowship
OpenAI on April 7, 2026 launched the OpenAI Safety Fellowship, a funded research program for external AI safety researchers. The pilot cohort runs September 14, 2026 through February 5, 2027. Fellows receive a weekly stipend of $3,850, approximately $15,000 per month in compute resources, and access to mentorship from OpenAI researchers. Applications close May 3, 2026, with notifications expected July 25. Research focus areas include safety evaluation, robustness, privacy-preserving methods, agentic oversight, and misuse mitigation. The program aims to expand the independent talent pipeline working on alignment and safety challenges as AI systems grow more capable.
Sources: OpenAI
Uber Expands AWS Contract to Pilot Amazon Trainium3, Pulling Back from Oracle and Google
Uber is expanding its Amazon Web Services contract to run ride-sharing features on Amazon’s in-house chips, announced April 7, 2026. The deal broadens Uber’s use of AWS Graviton ARM-based server processors and initiates a new pilot of Trainium3, Amazon’s Nvidia-competing AI accelerator chip. The move signals a strategic retreat from Uber’s 2023 multi-year commitments to Oracle Cloud Infrastructure and Google Cloud Platform. Uber joins Anthropic, OpenAI, and Apple as major companies that have expanded AWS usage for Amazon’s custom silicon. Amazon CEO Andy Jassy said Trainium was already a multibillion-dollar business by late 2025, underscoring growing competitive pressure on Nvidia in cloud AI compute.
Sources: TechCrunch
Visa Launches AI Agent Payment Platform With Trusted Protocol for Autonomous Commerce
Visa launched Intelligent Commerce on April 8, 2026, a platform enabling AI agents to make purchases autonomously across its payment network without traditional checkout flows. The system tokenizes card credentials, verifies purchases against user-defined spending limits, and routes transactions via Visa’s rails. A Trusted Agent Protocol authenticates AI shopping bots before granting transaction authority. Early testing partners include AWS, Aldar, Firmly, and Nekuda. General availability is planned for June 2026. Visa positions Intelligent Commerce as infrastructure for agentic e-commerce, which forecasters expect to exceed $5 trillion in transaction volume by 2030 as autonomous agents handle more consumer purchasing.
Sources: VentureBeat
NVIDIA Releases Isaac GR00T N1.5 and New Sim Tools During National Robotics Week
NVIDIA marked National Robotics Week—running April 4–12, 2026—with updates to its full physical AI stack. The company released Isaac GR00T N1.5, a new humanoid robot foundation model, alongside Isaac Sim 5.0 and Isaac Lab 2.2, its simulation and reinforcement-learning frameworks. Highlighted deployments include Aigen’s solar-powered weeding robots running on Jetson Orin hardware with NVIDIA Cosmos foundation models, and Maximo’s autonomous robot fleet completing a 100-megawatt solar installation. OpenClaw, an open-source robotic gripper, now runs inference locally on NVIDIA Jetson Thor for private, low-latency edge AI. Partner OEMs include ABB Robotics, FANUC, Agility Robotics, Figure AI, and Boston Dynamics.
Sources: The Robot Report
ICML Rejects 497 Papers After Watermarks Reveal Authors Used AI for Peer Reviews
The International Conference on Machine Learning rejected 497 submitted papers—roughly 2 percent of all submissions—after detecting that their authors used large language models to conduct peer reviews in direct violation of conference policy. Organizers uncovered the violations by embedding digital watermarks in research papers distributed to reviewers; any LLM-generated review reproduced the watermark text, exposing the fraud. ICML 2026, scheduled for Seoul in July, instituted two separate review streams: one permitting limited AI assistance and one prohibiting it entirely. The incident spotlights growing tensions in academic publishing over AI’s appropriate role in scientific evaluation and peer-review integrity.
Sources: Nature
Tech Pulse
Top Models: OpenAI o3 (71.7% SWE-Bench) | Gemini 2.5 Pro (63.8%) | Claude 3.7 Sonnet (62.3%)
Market: Apple $3.1T | Microsoft $2.9T | NVIDIA $2.8T
Curated by JD · samwise.agency