Anthropic's Mythos Preview has crossed a capability threshold in cybersecurity: it identified thousands of previously unknown zero-day vulnerabilities across every major operating system and browser, including a 27-year-old bug in OpenBSD and 271 Firefox flaws patched after a single evaluation pass. Access is restricted to a handpicked consortium of launch partners (AWS, Apple, Google, Microsoft, CrowdStrike, etc.) under Project Glasswing, while Anthropic withholds broader release due to offensive capability concerns. Over 40 additional critical infrastructure organizations have since been granted defensive access.
Euro-area finance ministers convened in Brussels to address a problem their regulations weren't designed for: no European bank, government, or cybersecurity agency has access to Mythos, while American tech firms use it to patch their systems. The European Commission confirmed it is in direct contact with Anthropic, and ministers are framing the access gap as a sovereign security risk, not merely a commercial grievance.
The Department of Defense finalized agreements with OpenAI, Google, Microsoft, Nvidia, AWS, Oracle, SpaceX, and Reflection to deploy AI on classified networks. Anthropic remains excluded after refusing to grant unrestricted access for autonomous weapons and mass surveillance. Pentagon CTO Emil Michael confirmed the blacklist stands despite White House-brokered meetings between Dario Amodei and senior Trump administration officials on April 17.
OpenAI closed its $10 billion joint venture, internally called "DeployCo," anchored by TPG with 19 investors including Brookfield, Advent, Bain, Dragoneer, and SoftBank. The structure guarantees investors 17.5% annual returns over five years. The vehicle's purpose: embed OpenAI products deep inside PE-owned portfolio companies, turning private equity into AI's primary distribution channel for enterprise adoption.
Announced the same day as OpenAI's DeployCo, Anthropic partnered with Blackstone, Goldman Sachs, Hellman & Friedman, General Atlantic, Apollo, and GIC to create an AI-native services firm. The venture will embed engineers inside mid-sized companies to redesign workflows around Claude agents, explicitly targeting the consulting industry's territory. Fortune characterized it as Anthropic "taking a shot at the consulting industry."
AI chip startup Cerebras commenced its roadshow, offering 28 million shares at $115–$125 each. The company reported $510M in 2025 revenue and holds a deal with OpenAI worth over $20 billion through 2030 for 2 gigawatts of compute capacity. TechCrunch called it "on track for a blockbuster IPO," marking the first major AI-native chipmaker to go public and a direct challenge to Nvidia's inference dominance.
Jensen Huang confirmed Nvidia's AI accelerator market share in China has fallen from 95% to zero, calling U.S. export policy something that "has already largely backfired." Nvidia took a $4.5B charge on stranded H20 inventory. Huawei, Cambricon, and Moore Threads are filling the gap, accelerating China's domestic AI chip ecosystem in a market projected to reach $67B by 2030.
Mayo Clinic's REDMOD system, validated in a landmark study published in Gut, detects pancreatic cancer on routine CT scans up to three years before clinical diagnosis — catching 73% of future cases versus 39% by radiologists. The system measures hundreds of quantitative imaging features capturing faint biological changes as cancer develops. Clinical trials (AI-PACED) are now underway for at-risk patients.
Google Mandiant's annual report, drawing on 450,000 hours of incident response, found that mean time-to-exploit has gone negative — 28.3% of CVEs are now exploited within 24 hours of disclosure. Threat actors are integrating LLMs as "force multipliers," with malware families like PROMPTFLUX querying language models mid-execution to evade detection. Attackers hand off compromised access in as little as 22 seconds.
OpenAI rolled out an opt-in "Advanced Account Security" mode for high-risk ChatGPT users — journalists, researchers, political dissidents — that mandates passkeys or hardware security keys, disables password and email recovery, shortens sessions, and automatically excludes conversations from training data. The company partnered with Yubico for bundled hardware key pricing. It's OpenAI's clearest acknowledgment that ChatGPT conversations are now sensitive enough to warrant nation-state-level protections.
