AI Daily 15 Apr: Claude Code Rebuilt, UK Flags AI Threats

The biggest AI story landing in developers’ hands today comes from Anthropic: the Claude Code desktop app has been rebuilt around parallel agent sessions, with a new automation layer called Routines that runs without a user present. Elsewhere on 15 April, OpenAI expanded access to a cybersecurity-focused variant of GPT-5.4, the UK’s AI Security Institute published concrete data on how far frontier AI has advanced as an attack tool, NVIDIA released the first open-source AI models for quantum computing, and Notion quietly became something closer to an application platform.

Anthropic shipped the most immediately usable developer update of the week: the Claude Code desktop app has been rebuilt around parallel agent sessions, giving technical teams the ability to run multiple tasks simultaneously without waiting for each to finish in sequence. The redesigned app introduces a new sidebar that puts every active and recent session in a single view. Users can filter by status, project, or environment, drag sessions to rearrange their workspace, and branch off a question from a running task using a keyboard shortcut without feeding extra context back into the main thread. The update also brings an integrated terminal, an in-app file editor, a rebuilt diff viewer for large changesets, and an expanded preview pane that handles HTML files and PDFs alongside local app servers.

The more consequential addition is Routines, launched in research preview. A Routine bundles a prompt, a repository, and any relevant connectors into a single configuration that can run on a schedule, fire from an API call, or trigger from a GitHub event such as a new pull request. The practical result is persistent automation: Claude working through a codebase overnight without anyone at the keyboard. Both parallel sessions and Routines are available now to Pro, Max, Team, and Enterprise subscribers via the redesigned desktop app or Claude Code on the web.

OpenAI launched GPT-5.4-Cyber on 14 April, a cybersecurity-tuned variant of its flagship model available only to vetted organisations through the company’s Trusted Access for Cyber programme. The model carries a calibration that reduces refusals on dual-use security queries, a friction point that frustrated professional users of earlier OpenAI models. Its headline capability is binary reverse engineering: analysing compiled software for malware or vulnerabilities without access to original source code, a task central to threat analysis and software supply chain security. Individual defenders can verify identity at chatgpt.com/cyber; enterprises can request team access through an account representative.

This is the second major restricted cybersecurity AI launch in April. Anthropic’s Claude Mythos Preview was distributed to 40 pre-vetted organisations earlier this month under similar logic: the capability exists, so better to shape how it reaches defenders than to leave security teams at a disadvantage while potential attackers find other routes.

The UK’s AI Security Institute published its evaluation of Claude Mythos Preview on 14 April, and the findings are specific enough to alter how regulators, banks, and security teams think about frontier AI capability. Mythos became the first AI model to complete a 32-step network attack chain from start to finish, with no human direction during the run. Claude Opus 4.6, the next-best model tested, averaged 16 of those 32 steps. On expert-level Capture the Flag challenges, tasks no AI could complete before April 2025, Mythos succeeded 73% of the time. The AISI noted that the tests used weakly-defended systems without live incident response, meaning these results describe capability against soft targets rather than hardened enterprise environments.

The publication is itself significant. The AISI does not routinely release granular attack-chain performance data for commercial AI models, and choosing to do so signals that the information is judged more useful in public hands than withheld. British financial regulators, including the FCA and Bank of England, are now in urgent conversations with banks about the risk profile of Mythos and models like it.

NVIDIA released Ising on 14 April, the first open-source family of AI models designed to make quantum computing practically usable at scale. Quantum processors have remained difficult to deploy reliably because of two persistent bottlenecks: calibration, which currently takes days of expert labour to tune a quantum chip, and error correction, which keeps calculations accurate during computation. The Ising Calibration model, a 35-billion-parameter vision-language model licensed under Apache-2.0, automates calibration and cuts setup time from days to hours. The Ising Decoding model handles real-time error correction at up to 2.5 times faster speed and three times better accuracy than existing tools. Early adopters include Fermi National Accelerator Laboratory, Harvard, IQM Quantum Computers, and the UK National Physical Laboratory.

Notion launched Workers for Agents in developer preview on 14 April, shifting the platform from a productivity tool into something closer to an application layer. The feature lets developers write custom server-side code that Notion’s built-in AI agents can invoke when responding to user prompts. Businesses can extend agent behaviour with their own logic rather than working within Notion’s native capabilities alone. The developer preview runs within a 30-second timeout and 128 MB of memory per execution, with opt-in through the developer portal. The update also introduces AI Autofill, which enriches and maintains database records automatically.

The Stanford Human-Centered AI Institute’s 2026 AI Index, released 13 April and covered extensively across technology media on 14-15 April, adds useful context for the pace of change. SWE-bench coding benchmark scores rose from 60% to near 100% in a single year. China has effectively matched the United States in frontier model performance, with labs in both countries now trading the leading position across benchmarks. AI model transparency is in decline: the Foundation Model Transparency Index averaged 40 points this year, down from 53, as the most capable models are increasingly closed. Generative AI reached 53% population adoption within three years, faster than either the personal computer or the internet.

At a Glance: Databricks brought Document Intelligence and Custom Agents to general availability under its new Agent Bricks platform, with an AI Gateway for enterprise governance. Jane Street committed $6 billion to CoreWeave’s AI cloud infrastructure, one of the largest single AI compute agreements by a financial firm. Meta’s Muse Spark, launched 8 April by the company’s new Superintelligence Labs and led by Alexandr Wang, is now live across Facebook, Instagram, WhatsApp, and Messenger in a closed-source configuration, a notable reversal of Meta’s previous open-source strategy. xAI’s Grok 4.20 Beta 2 continues to lead on major AI benchmarks, with Grok 5 now targeted for Q2 2026. Cursor 3, released 2 April, ships a proprietary frontier coding model running at over 200 tokens per second alongside a dedicated Agents Window.

Watch whether OpenAI expands TAC programme access in the coming days to UK and European cybersecurity organisations, which currently face a US-centred verification pathway. The FCA Supercharged Sandbox second cohort is the most concrete near-term UK regulatory catalyst, expected imminently following the March application deadline. Grok 5 is now targeting Q2 2026, making a May announcement realistic. And the UK King’s Speech on 13 May is the earliest point at which AI legislation could formally be introduced to Parliament.