Anthropic's Claude Mythos surfaces in product UI ahead of wider release

The model string claude-mythos-1-preview has appeared in Claude Code and Claude Security interfaces, and some users briefly saw a toggle to enable Mythos before it was pulled. Reporting ties the model to Project Glasswing, where it allegedly surfaced more than 10,000 high- and critical-severity zero-day vulnerabilities during early runs—Anthropic says it will release Mythos-class models more broadly once stronger safeguards exist.

OpenAI launches $4B Deployment Company, acquires Tomoro

OpenAI spun out a standalone enterprise unit—the OpenAI Deployment Company—backed by more than $4 billion from TPG, Brookfield, Advent, and Bain Capital, with a mandate to embed Forward Deployed Engineers directly inside customer organizations. The launch is paired with the acquisition of Tomoro, whose team has prior enterprise AI deployment experience at Tesco and Virgin Atlantic.

Anthropic ships 20-plus legal MCP connectors for Claude

Claude added more than 20 legal MCP connectors and 12 practice-area plugins covering research, contracts, discovery, matter management, and legal aid. The update expands Claude Cowork for law firms and in-house teams, and Anthropic says legal professionals have become the most engaged Claude Cowork users of any knowledge-work function since the first legal plugin launched earlier this year.

Meta's Muse Spark debuts as first proprietary, non-Llama flagship model

Meta unveiled Muse Spark, built under Chief AI Officer Alexandr Wang's newly formed Superintelligence Labs—a departure from the company's open-source Llama strategy. The model targets competitive performance on multimodal, reasoning, health, and agentic tasks at lower compute cost than Llama 4 mid-size, and is already powering Voice Mode and Meta Glasses in the US and Canada.

AISI Confirms Both Claude Mythos and GPT-5.5 Blew Past Cyber-Task Doubling-Rate Trends

The UK AI Security Institute published updated analysis showing that Claude Mythos Preview and GPT-5.5 both substantially exceeded the previously measured doubling rate for autonomous cyber-task completion — a rate that had already accelerated from 8 months per doubling in late 2024 to 4.7 months by February 2026. On a 32-step attack simulation, GPT-5.5 completed 2 of 10 runs versus Mythos's 3, putting a general-purpose model at near-parity with a restricted-access cyber-specialist. Whether this marks a new, faster trend or an isolated break is, per AISI, still unresolved.

Kimi K2.6 Hits 58.6% on SWE-bench Pro as Open-Weight MoE at Frontier Coding Level

Moonshot AI's Kimi K2.6 — a 1.6-trillion-parameter MoE model with 31 billion active parameters — scored 58.6% on SWE-bench Pro, matching GPT-5.5 and within 6 points of Claude Opus 4.7, while costing roughly 8x less per token at $0.60/$2.50 per million. The model's Agent Swarm architecture scales to 300 parallel sub-agents for long-horizon tasks; a 12-hour continuous tool-use trace porting an inference engine to Zig served as its public demo. The result, corroborated by the NIST CAISI cross-domain evaluation, marks the first open-weight model to reach the Western frontier ceiling on agentic coding tasks.

Charon simulator predicts LLM training and inference performance within 5% error

Charon is a unified, modular, fine-grained simulator that models large-scale LLM training and inference across parallelism strategies, system optimizations, and hardware configurations. It achieves overall prediction error consistently under 5.35%, and under 3.74% for large-scale GPU cluster training. In a practical inference deployment case, it discovered a configuration that improved system throughput over an engineering-tuned baseline.

Stanford HAI cuts LLM scaling-law prediction cost using measurement-science statistics

Researchers at Stanford HAI applied statistical concepts from measurement science and education to dramatically reduce the computational demand of predicting how the largest LLMs will scale. The method could save millions of dollars in training costs by making scaling predictions tractable without running full-scale experiments. The work was published May 21, 2026.

Illinois Senate Democrats race eight-bill AI package to May 31 deadline

With less than two weeks left in the spring session, Illinois Senate Democrats introduced an eight-bill package covering consumer protections, chatbot transparency, and restrictions on AI use in schools—explicitly citing the absence of federal action as motivation. Sponsors said that together with California and New York, the three states would cover roughly 40% of the U.S. AI market, signaling a coordinated multi-state approach to fill the federal vacuum. The package cleared committee with near-unanimous bipartisan votes and is now headed to full floor consideration before the May 31 adjournment.

New York Senate panel advances bill banning algorithmic price-setting by businesses

New York's S 8623 passed out of the Senate Consumer Protection Committee, moving a bill that would prohibit businesses from using algorithmically set prices and require disclosure of automated pricing systems. The measure mirrors Maryland's HB 895, which was signed into law in late April, making New York a potential second state to target AI-driven dynamic pricing in retail and services. If enacted, the bill would apply to any entity using an automated tool to determine prices for goods or services offered to New York consumers.