Good morning! Here's what's happening in AI today:
Anthropic reports three real-world security incidents from Claude evaluations
Google Earth adds Nano Banana AI image generation to real maps
Perplexity launches Projects, a persistent work hub inside Computer
How to build an AI marketing team inside Claude Code
4 new AI tools worth trying today
AI REPORT
Anthropic disclosed that a review of its cybersecurity evaluations turned up three incidents in which a Claude model reached the internet from inside a third-party evaluation environment and gained unauthorized access to real systems at three separate organizations, a review it ran jointly with evaluation partner Irregular.
The incidents happened while Claude was operating inside third-party evaluation environments, not in production use, but the model still found its way onto real infrastructure it wasn't meant to touch.
Anthropic is publishing exactly what happened and what it's changing as a result, and says it's encouraging other AI developers to run similar reviews of their own setups.
The review was done jointly with Irregular, one of Anthropic's evaluation partners, credited for the joint investigation.
If you run evaluations against Claude in a third-party environment, this is the incident report worth reading before your next test cycle. It's a rare case of a frontier lab publishing its own security failures in detail instead of quietly patching and moving on. For anyone building agentic systems that touch real infrastructure, the takeaway isn't that Claude is unsafe, it's that evaluation environments need the same access controls as production ones.
AI TOOL
Google Earth rolled out AI image generation powered by Nano Banana, letting users generate custom imagery directly on top of the platform's real satellite, aerial, and 3D data instead of a blank canvas.
You can now generate images anchored to actual places, using Google Earth's own satellite and 3D imagery as the base rather than working from a text prompt alone.
The demo reimagines a real site, Pompeii's amphitheater, with the generated layer sitting directly on the aerial view.
It works inside the existing Google Earth Pro interface, so there's no separate app or export step to get started.
This puts a real, image-grounded AI generator inside a tool millions of people already have open for maps and geography, no new account required. It's a meaningfully different use case than a blank-canvas generator, because everything you create is tied to an actual location. For anyone teaching, mapping, or building location-based content, that grounding is the entire point.
AI AGENT
Perplexity introduced Projects, an evolution of Spaces inside Perplexity Computer that gives users a persistent hub for ongoing work with a shared file system and memory that carries across sessions.
Projects replace the old Spaces model with something built for work that spans more than one sitting, so files and context stick around instead of resetting each session.
Perplexity frames this as a step toward a multiplayer agentic operating system, though the practical use today is simpler: one place to manage and collaborate on a task.
The feature is live for all users now, not gated behind a waitlist or a paid tier.
Most AI tools still treat every session like a blank slate, which means re-explaining context each time you return. Projects fixes that by keeping files, memory, and progress attached to the work itself, not the conversation. It's a real step toward Perplexity being useful for something longer than a single afternoon of research, and it ships free to every user today.
HOW TO AI
A real marketing team has a strategist, a copywriter, an SEO person, someone watching the competition. Most solo operators have themselves, a laptop, and however many tabs it takes to fake it.
There's an open-source Claude Code suite that gives you the team without the payroll. Five specialists, one orchestrator, free, MIT licensed, one line to install.
curl -fsSL https://raw.githubusercontent.com/zubair-trabzada/ai-marketing-claude/main/install.sh | bash
Method 1: Run the audit
Type /market audit <url> and the orchestrator launches five agents in parallel instead of running one script top to bottom. I will use a real use case below:
/market audit https://calendly.comCalendly came back at 74/100, Grade B, strongest in Growth & Strategy, weakest in Competitive Positioning at 60. Top quick win: the homepage headline reads as generic when it could lead with an actual authority claim already sitting in the trust badges.
Method 2: Rewrite the copy
Once the audit flags what's weak, /market copy <url> rewrites the actual page copy, before and after side by side.
/market copy https://calendly.com"Calendly makes scheduling simple" became "The scheduling platform 86% of the Fortune 500 run on," pulling a stat that was already sitting in the trust badges into the headline itself. It also caught four different CTA phrasings doing the same job and standardized them to one.
Method 3: Package it for a client
/market report-pdf <url> turns the same audit into something you'd actually hand over, cover page, score gauge, findings, competitor comparison.
/market report-pdf https://calendly.com
Diagnose, fix, deliver. Three commands, one afternoon, and the report that comes out the other end doesn't read like a chatbot summarized a URL. Read the full guide.
P.S. You can access all my prompts, workflows and AI trainings if you upgrade here
Cognition shipped native support for GitHub Stacked PRs in Devin, breaking large changes into smaller reviewable diffs and rebasing downstream changes automatically as the stack updates.
Figma added a new properties panel and annotations to Figma Make, so you can edit designs and prototypes built in code without leaving the canvas.
ChatGPT upgraded its Chrome extension with three changes: Side Chat can reference open tabs and YouTube videos, and the desktop app adds URL suggestions and browser history controls.
📱 Linear: Review agent code changes and leave line-specific feedback straight from the mobile app.
🔥 Firecrawl MCP: Agent-ready web scraping and search that now uses half the context of the previous version.
🎙️ ElevenLabs: Character Casting finds a distinct voice for every character in a manuscript before you generate the audiobook.
🎨 OpenAI Codex: Edit and refine images with a new lightbox and canvas view instead of rewriting the prompt each time.
THAT’S IT FOR TODAY
Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!
See you tomorrow :)
- Dr. Alvaro Cintas
✓ Full archive of premium guides with ready-to-use prompts
✓ Structured AI courses (step-by-step, start-to-finish)
✓ Every upcoming premium tutorial