AI brief — 19 May 2026

EU agrees to amend its AI Act, extending deadlines and banning nudifier apps

On 7 May, the European Parliament and Council reached a provisional agreement to reform the AI Act as part of the Digital Omnibus package. The deal extends compliance deadlines for most high-risk AI systems to December 2027 (for biometrics, critical infrastructure, education, employment, and law enforcement) and August 2028 (for AI embedded in products), giving organisations more time to prepare ahead of the original August 2026 cutoff. The amendment also introduces a new prohibition on AI systems that generate sexually explicit imagery from identifiable individuals — so-called "nudifier" apps — with compliance required from December 2026. Cyprus Deputy Minister for European Affairs Marilena Raouna welcomed the deal, saying it "supports our companies by reducing recurring administrative costs" and ensures "legal certainty" in implementing the rules. The formal vote by Parliament and Council is expected in the coming weeks.

Google opens I/O 2026 with Gemini 3.5, new pricing tiers, and a global coding agent

At its annual developer conference in Mountain View today, Google announced Gemini 3.5 Pro — the company's next-generation model — already in internal use, with wider availability expected in about a month. Gemini Spark, a lighter model, enters beta for AI Ultra subscribers next week. Google also reshuffled its subscription tiers: a new AI Ultra plan launches at $100 per month, while the existing Ultra plan drops from $250 to $200. Antigravity 2.0, Google's coding agent, is now available globally and was demonstrated autonomously building an operating system from scratch during the keynote.

US government completes pre-deployment safety pacts with all five major AI labs

The Center for AI Standards and Innovation (CAISI), part of the US Department of Commerce, confirmed on 5–6 May that all five major American frontier AI developers — OpenAI, Anthropic, Google DeepMind, Microsoft, and xAI — have signed voluntary pre-deployment evaluation agreements. Under the deals, companies submit versions of their models — sometimes with safety guardrails reduced — so CAISI can probe for national security risks before public launch. The agency's TRAINS Taskforce has completed more than 40 model evaluations to date, according to HPCwire. The agreements represent a notable shift for the Trump administration, which had previously resisted formal oversight structures for frontier AI.

Cursor ships Composer 2.5: frontier coding performance at a tenth of the price

Cursor released Composer 2.5 on 18 May, an AI coding model built on Moonshot AI's open-source Kimi K2.5 base and trained on 25 times more synthetic tasks than its predecessor. The model scores 79.8% on SWE-Bench Multilingual — comparable to Anthropic's Opus 4.7 and OpenAI's GPT-5.5 — at a price of $0.50 per million input tokens, roughly one-tenth the cost of either flagship model. Cursor also disclosed that it is training a substantially larger model from scratch in partnership with xAI, using Colossus 2's million H100-equivalent accelerators.

Anthropic's Mythos model found thousands of unknown software vulnerabilities — and is keeping them under lock

Anthropic's unreleased Claude Mythos Preview has identified thousands of critical zero-day vulnerabilities across every major operating system and web browser, including a 17-year-old remote code execution flaw in FreeBSD (CVE-2026-4747) that could give an attacker complete control of a server. Rather than releasing the model publicly, Anthropic is distributing access through Project Glasswing, a controlled consortium that includes AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, and Nvidia. The aim is to give defenders time to patch critical systems before models of similar capability become broadly available, according to Fortune.

Anthropic in talks to raise $30 billion at a valuation above $900 billion

Anthropic is in discussions to raise at least $30 billion in new financing at a pre-money valuation exceeding $900 billion — which would for the first time surpass OpenAI's $852 billion valuation from March 2026, according to Bloomberg. The round is expected to be co-led by Sequoia Capital, Dragoneer, Greenoaks, and Altimeter Capital, with closing expected before the end of May. For context, Anthropic raised $30 billion in a Series G round as recently as February 2026 at a $380 billion post-money valuation; the new talks reflect how quickly investor appetite has grown alongside the company's revenue and model capabilities, also reported by CNBC.