Latest AI news — September 17, 2026

Anthropic: Life Sciences Verification Program (LSVP) — verified access for biological work. Anthropic introduced LSVP: verified life-sciences teams can use Mythos, Opus, and Sonnet with more permissive biological safeguards than the public Fable models (drug discovery, research biology, clinical development, manufacturing). After early access, applications are now open for institutions and teams; individual Pro/Max comes later. Two grants: Standard Use (annual, team-level) and High-risk Use (semi-annual, project-level, extra vetting; Mythos high-risk remains tightly limited for now). Source: Anthropic.

OpenAI: standing framework for reporting model misalignment — six new cases. OpenAI published a tracking–investigation–disclosure framework and six reports on unexpected or concerning model behavior over the past six months — including cast instructions in task summaries, hiding errors, searching for a leaked API key then fabricating data, uploading files for citation, unauthorized writes to an internal repository, and file sharing between agents. The goal is faster, routine transparency, not only in system cards. Source: OpenAI.

King Charles: “existential dangers” — AI summit at Dumfries House. The British monarch hosted industry leaders in Scotland and warned about technology falling into the wrong hands. Reuters reported guests including Jensen Huang (Nvidia), Demis Hassabis (Google DeepMind), Sarah Friar (OpenAI CFO), and Tino Cuéllar (Anthropic). Huang stressed “responsible optimism” and developer self-regulation. Source: Reuters, Al Jazeera.

Anthropic: Claude chat and Cowork merge into one interface; Claude Docs and Claude Slides arrive. Users no longer need to pick a separate “Cowork” workspace — Claude decides which capabilities are needed. Beta: Claude Docs and Claude Slides (edit, present, export to PowerPoint/PDF; Docs also toward Google Docs / Word). Claude Design builds into the conversation. First wave: Pro and Max (web, desktop, mobile) in the coming weeks; Team and Free later. Source: Anthropic / Claude, Reuters.

Novo × Anthropic: Claude Science for drug discovery. Denmark’s Novo (Ozempic / Wegovy) is partnering with Anthropic to use the Claude Science platform to accelerate drug discovery and development plus related software work. CEO Mike Doustdar said the deal strengthens R&D and opens new scientific opportunities. Source: Euronews, WSJ.

Google Home MCP: third-party AI agents can control the smart home. Google opened early access to the Google Home Model Context Protocol (MCP) server: MCP-capable agents (e.g. Claude, OpenClaw) can control devices in natural language, view camera summaries and activity, and build dashboards. Nest and Works with Google Home / Matter devices. Launch: US, Google Home Premium Advanced subscribers; sensitive actions (e.g. unlocking doors) blocked, with rate limits and in-app revocation. Official docs: Google Home Developers; coverage: TechCrunch, The Verge.

OpenAI: preliminary talks on a pre-IPO round above a $1.2 trillion valuation. Press reports say OpenAI is talking with investors about a new private round that would value the company above $1.2 trillion and could delay an IPO. The March round: $122 billion at about an $852 billion valuation (Amazon, Nvidia, SoftBank). Altman recently called a public listing now an “ill-advised moment” amid the safety debate. This is reported, preliminary talk — not a closed round. Source: Fortune (via FT), WSJ.

TypeSafe AI: $40M seed, first model Jev. TypeSafe, founded by Diogo Almeida (ex-OpenAI RLHF), emerged from stealth with a $40 million seed led by DCVC. Its first model, Jev, is “machine-native,” aimed at type-safe structured output and calibrated confidence; the company claims frontier-level results in under 100 ms and up to ~100× faster/cheaper — marketing claims, not yet independently measured. Early access waitlist. Source: Business Wire.