AI News: June 27, 2026

1. Trump Administration Clears Anthropic’s Mythos for 100+ Agencies and Companies

The Trump administration authorized Anthropic to make its Mythos 5 model available to more than 100 specific U.S. government agencies and companies, with Commerce Secretary Howard Lutnick reversing a roughly two-week ban on the cybersecurity-capable model. The directive permits non-American employees at those trusted organizations, including Anthropic’s own international staff, to access Mythos 5, while the more widely available Fable 5 variant remains unreleased for general use. The customer-by-customer clearance signals that frontier model distribution is now subject to government gatekeeping, a constraint practitioners will have to plan around when sourcing high-capability models. Source

2. Analysis: The Real AI Contest Is Now Labs Versus Regulators

TechCrunch AI editor Russell Brandom argued that the rivalry between Anthropic and OpenAI has become secondary to a shared problem: government control over whether frontier models can ship at all. Both companies now face customer-by-customer approval regimes, with Anthropic’s Mythos stuck in preview for months and OpenAI’s GPT-5.6 rollout gated by federal sign-off, raising questions about how labs recoup training costs when release timing is out of their hands. The piece contends regulators lack the technical expertise to define what safety assurances would even satisfy them, leaving the industry without a clear path to predictable launches. Source

3. OpenAI’s IPO May Slip to 2027 as Altman Holds Out for $1 Trillion

OpenAI is reportedly postponing its public listing to 2027 because CEO Sam Altman is insisting on a $1 trillion valuation and rejecting anything lower, according to New York Times reporting. The company’s last private valuation was $730 billion, it remains unprofitable on roughly $13 billion in 2025 revenue, and ChatGPT user growth has stalled near 900 million, short of its billion-user target. Advisers also cited volatile tech markets and SpaceX’s weak post-IPO performance as reasons to wait, underscoring how even the most prominent AI lab faces skepticism about translating scale into sustainable economics. Source

4. Linux Foundation and 20 Tech Firms Launch Akrites to Patch Open Source Before AI Attacks

The Linux Foundation launched Akrites, a coordinated initiative with roughly 20 founding members including AWS, Anthropic, Google, IBM, Microsoft, NVIDIA, OpenAI, and Red Hat, to find and fix vulnerabilities in widely used open source software before AI tools can exploit them. The effort sets up a single Security Incident Response Team to vet reports, filter duplicates, and coordinate confidential fixes, responding to the reality that AI models can now scan large codebases for flaws in minutes. The group notes that fewer than five percent of thousands of validated open source vulnerabilities from recent months have been patched, a gap that becomes far more dangerous as automated attackers accelerate. Source

5. MirrorCode Benchmark Shows a Model Coding for 19 Days Straight on One Task

Epoch AI and METR released MirrorCode, a benchmark that asks models to reimplement complete programs from scratch without access to the original source, where one large task ran an AI model continuously for 19 days at a cost of $2,600 with zero human intervention. Claude Opus 4.7 led with a 56 percent overall solve rate and rebuilt a 16,000-line Go bioinformatics toolkit in about 14 hours, work estimated to take humans two to seventeen weeks, though no model solved any of the largest tasks. The results quantify both how far sustained autonomous coding has come, up from roughly 30 percent a year earlier, and where it still breaks down on the hardest long-horizon work. Source

6. Lindy Drops Claude for Deepseek, Saving Millions as Cost Pressure Builds

Lindy, a 25-person AI agent startup, switched entirely from Anthropic’s Claude to Deepseek after model costs grew unsustainable and exceeded the company’s personnel expenses, with CEO Flo Crivello saying the move saved millions. Crivello described the expense curve as having “crashed to the ground” but said he would return to Claude if Anthropic cut pricing, framing the decision as economic rather than a verdict on quality. The switch is a concrete data point in a broader trend of teams adopting cheaper Chinese models like Deepseek and GLM-5.2 for agentic workloads, where high token volumes make per-token price a decisive factor. Source

7. Insurers Turn to Diffusion Models for Catastrophe Risk, but Hallucinations Loom

The insurance industry is adopting generative AI to model catastrophe risk, with firms including Fathom, Verisk, and Moody’s RMS using diffusion models to synthesize tens of thousands of weather scenarios beyond what historical data supports, the Financial Times reported. Fathom trained its models on roughly 1,000 years of climate simulations and refined precipitation resolution from 100km to 10km grids, but its scientific director warned the models can produce events that look plausible while violating basic physics. A deeper concern is incentive alignment, since insurers may favor models that yield lower loss estimates and let them write more business even when real risk is higher. Source

8. Report: Over Half of Grok’s Traffic Tied to Adult Content

xAI’s Grok sees well over half of its traffic going to pornographic images, videos, roleplay chats, and other adult content, according to two former employees cited by The Information. SpaceX IPO filings indicate Grok generated roughly 10 billion images and 2 billion videos per month in the first quarter of 2026, and the report says xAI delayed acting against non-consensual intimate imagery until regulatory pressure forced intervention. The account, which notes all of Grok’s co-founders have since departed, highlights the content-moderation and reputational exposure that comes with deploying loosely governed generative media at scale. Source