Today's stories. Each one has an "Ask Claude about this" button at the bottom — live follow-ups in your own words.
i.
Research · Top Story
Claude Mythos Delivers the Cybersecurity Reckoning That Everyone Saw Coming
Last week
Anthropic's Claude Mythos Preview is finding zero-day vulnerabilities faster than humans can process them, identifying thousands of critical flaws across every major operating system and web browser. The model discovered a 27-year-old vulnerability in OpenBSD and a 16-year-old bug in FFmpeg that millions of automated security scans had missed completely.
The model doesn't just find vulnerabilities—it builds working exploits autonomously, completing attack chains in hours that expert penetration testers said would take weeks. Internal testing showed Mythos could identify and exploit vulnerabilities in over 83% of attempts, fundamentally changing the economics of offensive cybersecurity capabilities.
"AI capabilities have crossed a threshold that fundamentally changes the urgency required to protect critical infrastructure from cyber threats, and there is no going back."
— Microsoft partner on Mythos
Anthropic responded by restricting access through Project Glasswing, a consortium including Apple, Microsoft, and Amazon focused on defensive applications. The move triggered immediate policy discussions in Washington and prompted OpenAI to accelerate its own GPT-5.5-Cyber rollout to vetted security teams. The message is clear: the age of AI-driven vulnerability discovery has arrived, whether defenders are ready or not.
Google is in talks with SpaceX to launch orbital data centers as part of Project Suncatcher, bringing AI compute infrastructure to space where solar power is constant and local backlash is impossible. The discussions come as SpaceX prepares its $1.75 trillion IPO, with orbital data centers positioned as a key revenue driver alongside Starlink.
Anthropic has already committed to using SpaceX's Memphis data center and expressed interest in future orbital facilities, while Google plans prototype satellite launches by 2027. The logic is compelling: space offers unlimited solar power, no cooling requirements, and freedom from terrestrial regulatory constraints that increasingly hamper ground-based buildouts.
The economics remain questionable in the near term. Today's terrestrial data centers are significantly cheaper than orbital alternatives when satellite construction and launch costs are factored in. But advocates argue that declining launch costs and increasing local opposition to ground-based facilities could flip that equation within years, not decades.
General Motors laid off approximately 600 IT workers this week while simultaneously hiring AI-native engineers, data scientists, and autonomous vehicle specialists. The cuts represent over 10% of GM's IT workforce, primarily affecting employees in Austin and Warren who lack skills in agent development, model engineering, and cloud workflows.
This isn't permanent downsizing—it's skills arbitrage. GM is actively recruiting for AI-focused roles including prompt engineering, agentic workflow development, and neural network architecture. The company hired Behrad Toghi from Apple as AI lead and Rashed Haq as VP of autonomous vehicles, signaling a fundamental shift in technical priorities.
"GM is transforming its Information Technology organization to better position the company for the future."
— GM statement
The move reflects broader industry trends where AI literacy has become mandatory rather than optional. Companies are finding it more efficient to swap entire teams than retrain existing staff, creating a two-tier job market split between AI-enabled roles and everything else.
South Korea's AI Tax Talk Tanks Semiconductor Stocks
Yesterday
South Korean presidential policy chief Kim Yong-beom proposed redistributing AI profits to citizens through a national dividend, triggering a 5.1% intraday drop in the Kospi index before markets recovered. Samsung fell 2.3% and SK Hynix dropped 2.4% as investors interpreted the comments as potential windfall tax legislation targeting semiconductor giants.
The proposal reflects growing political pressure as Samsung is forecast to generate 330 trillion won in operating profit this year, potentially ranking second only to Nvidia globally. Kim argued that AI gains built on decades of national infrastructure investment should be shared beyond corporate boardrooms, citing the sacrifice of farmers and fishermen who subsidized the tech buildout.
President Lee Jae Myung later clarified the comments referred to excess tax revenue from AI boom profits, not direct corporate taxation, but the damage was done. The episode revealed how quickly redistribution rhetoric can destabilize markets that depend on concentrated AI windfalls—a preview of political tensions as AI profits become too large to ignore.
DeepSeek's V4 model series represents the most capable Chinese AI system to date, with the V4-Pro achieving a Codeforces rating of 3206—making it the first open-source model to match closed-source performance on competitive programming. The model features 1.6 trillion parameters with only 49 billion activated per token, dramatically improving cost efficiency.
However, NIST evaluations suggest V4's capabilities lag the US frontier by approximately 8 months, performing similarly to GPT-5 rather than current-generation models like GPT-5.5 or Claude Mythos. DeepSeek acknowledges this gap explicitly, stating V4 trails state-of-the-art frontier models by 3 to 6 months on reasoning and knowledge benchmarks.
"DeepSeek's V4 preview is a serious flex, offering lower inference costs than previous models."
— Neil Shah, Counterpoint Research
The model's optimization for Huawei's Ascend chips rather than Nvidia hardware signals Beijing's push for semiconductor independence, but the performance tradeoffs remain significant. V4 proves Chinese AI capabilities are advancing rapidly while confirming the US maintains meaningful technological advantage.
OpenAI released GPT-5.5-Cyber to vetted security teams just days after Anthropic's Mythos announcement, offering reduced safeguards for authorized defensive workflows. The model enables vulnerability research, exploit development, and red team automation for approved organizations through the Trusted Access for Cyber program.
The timing wasn't coincidental. While Anthropic restricted Mythos to 40 organizations, OpenAI is taking a more open approach with tiered access levels based on identity verification and organizational vetting. Both companies are racing to arm defenders before offensive capabilities proliferate to adversaries.
Rising oil prices and technology stock weakness halted Wall Street's record-breaking run, with the S&P 500 dropping 0.4% from all-time highs. The selloff began in Asian markets where South Korea's Kospi tumbled on AI profit redistribution concerns, spreading to US tech giants as investors questioned sustainability of AI valuations amid energy cost pressures.
The Department of Energy proposed $1.2 billion for three new AI supercomputers at national laboratories, separate from the agency's main budget that faces a 13% cut. The investment signals federal prioritization of sovereign AI infrastructure even as biological and environmental research programs face 54% funding reductions.
Australian organizations are expected to spend $33.6 billion on public cloud services in 2026, up 17.9% year-over-year, driven primarily by AI workload demands. Infrastructure-as-a-service spending leads growth at 24.1% annually as companies rush to secure compute capacity for AI applications, reflecting global competition for cloud resources.