The AI briefing for people who are tired of AI briefings.

Breaking News

OpenAI Built a Cyber AI for Defenders

OpenAI released GPT-5.4-Cyber, a new model designed to help security professionals combat threats, but only vetted defenders can use it.

Why it matters: AI tools are getting better at finding software flaws. The race is on to put those tools in the hands of the good guys before the bad guys catch up. OpenAI is betting that giving defenders stronger, less restricted AI (under tight controls) is safer than holding back.

The details: GPT-5.4-Cyber is a fine-tuned version of GPT-5.4 with fewer guardrails for security tasks. It can reverse-engineer compiled software to hunt for malware or vulnerabilities. Access is gated through OpenAI's Trusted Access for Cyber program; higher trust levels unlock more powerful features.

By the numbers: OpenAI says its broader cyber tools have contributed to over 3,000 critical security fixes and scanned more than 1,000 open-source projects. Those numbers are self-reported, not independently verified.

The big picture: This launch came one week after Anthropic unveiled Project Glasswing and its Mythos model for cybersecurity. Both companies are restricting access to their most powerful cyber tools. Restricted-access cyber AI is becoming a product category of its own.

Yes, but: More permissive AI for security work is a double-edged sword. The same tools that help defenders find vulnerabilities could help attackers exploit them. The real test will be how well those controls hold up at scale.

What to watch: OpenAI says even more capable models are coming in the next few months. Anthropic plans to publish lessons from Project Glasswing within 90 days. How both companies handle the next wave of stronger cyber AI will shape the rules for the entire industry.

Labs

Claude Code Now Runs While You Sleep

Anthropic just gave its AI coding tool the ability to do your repetitive dev work on autopilot, no terminal required.

Why it matters: Until now, using Claude Code meant sitting in front of a screen and guiding the AI through tasks. Routines flip that model. Developers can now save a task once and have it run on a schedule, from an API call, or when something happens in GitHub. The work runs on Anthropic's cloud, even if your laptop is off.

Think of it as turning Claude Code from a hands on assistant into a set it and forget it coworker.

The details: A routine packages three things: a prompt (what you want done), one or more repos (where to do it), and connectors (tools like Slack, Linear, or Google Drive). You pick a trigger and Claude Code handles the rest in a full cloud session.

Three trigger types are supported at launch:

• Schedules: Hourly, daily, weekly, or custom cron expressions.
• API calls: Each routine gets its own HTTP endpoint and bearer token, so external tools can kick off a run.
• GitHub events: Pull requests and releases can automatically start a session. Filters let you target specific branches, authors, or labels.

Anthropic says the sweet spot is clear, repeatable work: things like PR review, backlog grooming, deploy checks, doc updates, and alert triage.

By the numbers:

• Pro plan: 5 routine runs per day
• Max plan: 15 runs per day
• Team and Enterprise: 25 runs per day
• Each cloud session gets roughly 4 vCPUs, 16 GB of RAM, and 30 GB of disk

How it fits in: Routines are not meant to replace everything. Anthropic now frames four automation lanes:

• Routines for unattended cloud jobs on triggers.
• Desktop scheduled tasks when you need local files or tools.
• GitHub Actions when you want CI that lives in your repo config.
• /loop for temporary tasks tied to an open CLI session.

Routines fill the gap where developers were duct taping together cron jobs, shell scripts, and CI pipelines to get Claude Code to run on its own.

Yes, but: This is still a research preview. Anthropic says the API, limits, and behavior can all change. There is no dedicated secrets store yet for cloud sessions, so environment variables are visible to anyone who can edit the environment. And routines belong to individual accounts, not teams, which could get messy when organizations want shared, centrally owned automation.

There is also a catch for enterprise customers using Zero Data Retention: since routines are part of Claude Code on the web and require server side storage, ZDR orgs likely cannot use them yet.

What to watch: Whether Anthropic moves routines out of research preview and adds team ownership, a proper secrets manager, and support for code forges beyond GitHub. That would signal a serious push to make Claude Code the default automation layer for software teams.

Product

Chrome Turns Prompts Into Reusable Tools

Google is embedding saved AI workflows directly into Chrome, letting users turn their best prompts into one-click commands that work on any page they browse.

Context: Over the past year, Google has built an AI assistant directly into Chrome called Gemini. It lives in a side panel next to whatever page you're browsing. Unlike going to a separate AI website, Gemini in Chrome can actually see the page you're on. It can read your current tab, pull in up to 10 open tabs at once, and connect to Google apps like Gmail and Drive.

Think of it like having a research assistant sitting next to you while you browse. You can ask it to summarize an article, compare products across tabs, or draft an email based on what you're reading.

Why it matters: Until now, every prompt you typed into Gemini in Chrome was a one-time thing. You'd craft a great prompt, get a useful answer, and then lose that prompt forever. Google's new "Skills" feature changes that. It lets you save your best prompts and reuse them like mini apps inside your browser.

How it works: Skills builds on top of Gemini in Chrome's ability to read your open tabs.

When you write a prompt that works well, you can save it as a Skill. Next time you need it, just type / or click + to pull it up and run it on whatever page you're viewing.

Google is also shipping a library of premade Skills for common tasks like comparing product specs across tabs, scanning long documents, or breaking down ingredient lists.

Not a full AI agent: Skills is not the same as Chrome's "auto browse" feature, which can take over your browser and complete multi-step tasks across websites. Skills is lighter than that. It replays a saved prompt with fresh page context. Think of it as a recipe you can rerun, not a robot that drives your browser.

Who gets it: Skills is rolling out now on desktop (Mac, Windows, ChromeOS) for users in the U.S. with Chrome set to English. Your saved Skills sync across any signed-in Chrome device.

Gemini in Chrome itself is only available in the U.S., Canada, India, and New Zealand, so the broader feature set still has geographic limits.

The bigger picture: Google has been layering AI into Chrome for months: first a chat assistant, then connections to Google apps, then auto browse for hands-free tasks. Skills fills the gap between chatting with AI and handing it full control. It's the "save and reuse" layer that makes casual prompting feel more like a real workflow.

Yes, but: Saved prompts can be more sensitive than one-off questions, especially at work. Google says enterprise accounts get extra protections: prompts and page data won't train AI models and won't be reviewed by humans outside the organization.

What to watch: Google launched Skills in English only and on desktop only. How fast it expands to other languages, mobile, and markets will signal whether Google sees this as a core browser feature or a niche add-on.

Policy

UK Tells Firms: AI Makes Hackers Faster

The UK government warned businesses of all sizes that AI models can now find and exploit software flaws at a speed that was unthinkable a year ago, and told them to shore up defenses now.

Why it matters: This is not just another cyber warning. For the first time, UK ministers moved past cautious forecasts and issued a direct, economy wide call to action aimed at every business, not just critical infrastructure. The message: if your defenses are weak, AI will help attackers find out faster.

What happened: The open letter, published April 15, called out Anthropic's Mythos model by name. The UK's AI Safety Institute tested it and said it is "substantially more capable at cyber offence than any model we have previously assessed."

In controlled tests, Mythos completed a full 32 step simulated corporate network attack, something no model had done before. It solved expert level capture the flag challenges 73% of the time.

How fast is this moving: AISI now estimates that AI cyber capabilities are doubling every four months. That is twice as fast as the eight month estimate it published just five months earlier. In early 2024, frontier models could handle basic cyber tasks about 10% of the time. Now the best models clear expert level tasks most of the time.

The real target: The government is not saying AI can already break into well defended systems on its own. The tests ran on networks with no active defenders and no alerts. AISI itself said it cannot conclude Mythos would succeed against hardened targets.

The real warning is simpler: businesses with slow patch cycles, poor access controls, thin monitoring, and old systems will be hit first and hardest as these tools spread.

What the government wants you to do: The advice is surprisingly basic. Ministers did not push exotic AI defense products. They said:

• Treat cyber risk as a board level issue using the Cyber Governance Code of Practice.
• Get certified under Cyber Essentials.
• Sign up for NCSC Early Warning alerts.

The logic: AI powered attacks will punish the same old gaps that already cost UK businesses an estimated £14.7 billion a year. 43% of businesses reported a breach in the past 12 months before this new wave even hits.

Yes, but... Mythos is not publicly available. Anthropic is keeping it under tight, limited release for defensive use only. So the direct risk from this specific model is low for now. But the NCSC's view is that AI cyber tools will spread across commercial and open source systems, broadening access to offensive capability over time.

What to watch: Keep an eye on whether AISI publishes the full technical basis for its new four month doubling estimate. The benchmark results are public, but the revised timeline has not been fully unpacked yet. If that pace holds, the gap between what attackers can do and what most businesses can defend will widen fast.

Investment

Jane Street Bets $7 Billion on AI Compute

Trading giant Jane Street is spending roughly $7 billion to lock down access to AI computing power, striking a massive deal with cloud provider CoreWeave that includes a $6 billion compute contract and a $1 billion equity stake.

Why it matters: GPU computing power, the fuel that trains AI models, has become one of the scarcest resources in tech. Companies across industries are now willing to spend billions and sign years long contracts just to guarantee they can get access to it. The era of pay as you go cloud computing is giving way to something that looks more like an arms race.

The deal: Jane Street will get access to CoreWeave's GPU clusters across multiple data centers, including next generation Nvidia chips. The $6 billion compute contract is one of the largest ever signed by a single company.

On top of that, Jane Street bought $1 billion worth of CoreWeave stock at $109 per share, about 7% below market price. That makes the trading firm one of CoreWeave's five largest shareholders.

Not alone: This is the third mega deal CoreWeave announced in a single week. It also expanded a $21 billion infrastructure agreement with Meta and signed a multi year deal with Anthropic. CoreWeave's contracts with OpenAI alone total about $22.4 billion.

The bigger picture: Demand for AI compute is growing faster than anyone can build it. Nvidia says it is making chips as fast as its factories allow, and still cannot keep up. GPU rental prices jumped over 20% in just three months this year, even for older chip models.

Amazon's CEO has said that some customers tried to reserve AWS's entire planned compute capacity for the coming year. Data center projects are being delayed by power shortages and permit backlogs. As one analyst put it: "Everything is sold out across the board."

Yes, but: CoreWeave is spending big to meet this demand. It plans $30 to $35 billion in capital spending this year alone, more than double last year. Its long term debt now tops $14 billion. If AI demand ever cools, that debt load could become a serious problem.

What to watch: Whether other industries beyond tech and finance start making similar billion dollar compute commitments. If they do, the supply crunch could get even tighter, and the companies that locked in capacity early will hold a major advantage.

Research

Google's New AI Voice Sounds Scarily Real

Google just launched Gemini 3.1 Flash TTS, a text-to-speech model that lets anyone direct AI voices like a movie director, at a fraction of the usual cost.

Why it matters: AI voices have long sounded flat and robotic. This model changes the game by letting users control tone, emotion, and pacing with simple text commands. Think of it as going from reading a script to actually performing it.

How it works: The model uses over 200 "audio tags," simple keywords you drop into your text to shape how the voice sounds. Want the AI to whisper? Add [whispers]. Need it to sound excited? Add [excited]. You can even mix tags to get exactly the delivery you want.

For example, you could write:

> [determined] We must save the mission. [short pause] No one else can! [excitement]

And the AI will shift its voice to match each cue, like an actor following stage directions.

By the numbers: The model scored an Elo of 1,211 on a major independent benchmark, placing it among the best in the industry for natural sound. It costs about $30 per million characters of text, roughly 5x cheaper than a similar ElevenLabs model. It supports 30 built in voices and over 70 languages.

Go deeper:

Multi-speaker dialogue. The model can handle two or more AI voices talking back and forth in one audio file. No more stitching clips together. This makes it useful for podcasts, audiobooks with multiple characters, or customer service bots that hand off between voices.

Built in watermarks. Every clip made with this model carries a hidden SynthID watermark. You can't hear it, but detection tools can spot it. This helps flag deepfakes and AI generated audio, a growing concern as voice cloning gets better.

Wide access. Developers can use it through Google's Gemini API, AI Studio, or Vertex AI. Google Workspace users can also tap into it through Google Vids, a no code video tool.

Yes, but... The audio tags only work in English for now. You can still generate speech in 70+ languages, but you have to write your control instructions in English. And some rival models still edge it out in raw realism. Google's real advantage here is the mix of low price, high quality, and deep control, not pure sound perfection.

What to watch: If Google keeps this pricing and quality, expect competitors like ElevenLabs and OpenAI to respond fast. The bigger shift is the "director style" control model. If this catches on, it could reshape how podcasts, ads, and audiobooks get made, with far smaller teams and budgets.