Good morning! Here's what's happening in AI today:
Claude's share feature exposed sensitive user data through Google search
OpenAI's ChatGPT Work agent can now log into sites that require sign in
Gemini Spark widens to Google AI Pro subscribers in the US
How to get better results from Claude without adding more prompts
4 new AI tools worth trying today
AI NEWS
IBTimes UK reported that hundreds of publicly shared Claude conversations, including some containing API keys, crypto wallet details, and resumes, turned up in Google search results this week, after Om Patel - an X user flagged the issue on the platform.
Claude's share feature builds a public web page any time you click Share, and that page stays live even after the conversation continues, only the snapshot at share time is exposed, not later messages.
Anthropic's robots.txt already instructs crawlers to skip these pages, but once a link gets posted anywhere public, a forum, a blog, a tweet, search engines can find it through that external reference instead.
Some developers pushed back on the framing, pointing out these weren't private conversations that leaked, they were links users chose to make public. Anthropic hasn't commented publicly yet, and Google had already stopped returning many of the pages by Thursday.
If you've ever used Claude's share button, especially for code, financial questions, or anything with a real name attached, it's worth five minutes to check. Go to Settings, then Privacy, then Shared Chats, and revoke anything you don't want sitting in a search index. Team and Enterprise accounts were never affected, this only touches Free, Pro, and Max.
AI AGENT
OpenAI announced that ChatGPT Work agents can now access websites that require a sign in, and your login persists across sessions so you only have to enter it once.
You take over the cloud browser to log in yourself, then hand control back to the agent to keep working, so credentials never get typed by the model itself.
This unlocks a category of tasks that were previously off limits to agents entirely, checking a loyalty account, booking through a site that requires an account, or pulling data from an internal tool behind a login wall.
The persistence matters more than the login itself. Without it, every single task would mean logging in again, which defeats the point of handing work off in the first place.
An agent that can't get past a login screen can't actually finish most tasks, it can look up a reservation but not book it. Persistent login is what closes that gap. Worth watching what happens if a saved session ends up somewhere it shouldn't.
AI PRODUCTIVITY
Google confirmed that Gemini Spark, its background agent that works 24/7 on tasks under your direction, is now rolling out to Google AI Pro subscribers in the US, after launching to Ultra subscribers earlier this month.
Spark runs independently of an active chat session, meaning it can keep working on a task after you close the app, rather than only acting while you're present.
The Pro rollout is US-only for now, with Google saying more countries and languages are coming, following the same staggered pattern Ultra got two weeks earlier.
This puts a persistent background agent in front of a much larger subscriber base than the Ultra-only release did, Pro is Google's more common paid tier.
A background agent only earns its keep if enough people can reach it, and Ultra was always a small, expensive tier to judge that on. Pro is Google's actual mainstream plan, so this is the real test of whether Spark holds up once it's running for millions of accounts instead of a few. If it works here, expect it in the free tier before long.
HOW TO AI
Anthropic's Claude Code team cut over 80% out of their own system prompt for Opus 5 and Fable 5. Then they re-ran the coding evals. No measurable loss.
The problem was contradiction. "Leave documentation as appropriate" sitting next to "do not add comments." Both live in the same context window, both technically rules the model has to follow. Older models needed a hard floor like that, they'd guess wrong too often without one. Opus 5 and Fable 5 don't fail that way anymore. So the floor just blocks good outcomes now instead of preventing bad ones.
Four things to go do with your file this afternoon.
Cut the "never" list
Anthropic's old line banned multi-paragraph docstrings, banned planning docs, banned anything not explicitly requested. Falls apart the second someone actually wants documentation. The replacement is one sentence: write code that reads like the surrounding code. Most rewrites end up a third to a tenth of the original length. Nobody needs zero lines in their file. They need it to stop repeating itself.
Split one bloated skill file into a tree
Code review, testing, deployment. If all three live in one file and load into every session no matter what the task is, that's the fix:
.claude/skills/
code-review.md
test-writing.md
deployment-verification.mdA bug fix never touches deployment-verification.md now. Name the files for the job, not the tool underneath, test-writing.md instead of pytest.md. Tools get swapped out. The job stays the same.
Redesign the tool instead of explaining it
Examples narrow what the model tries, it pattern-matches to the example even when a different approach would've worked better. A status field written as a plain string invites guessing. The same field as an enum, pending, in_progress, completed, teaches itself. Add one line on top, keep exactly one item in progress at a time, and you're done. No example required.
Let memory handle itself
The # hotkey used to be how you told Claude to remember something. Not broken now, just redundant. Run a session without touching it once, then check Settings, then Memory. Most of what you'd normally save by hand is already sitting there. If something's genuinely missing, that's not a reason to go back to manual saves. It means that fact belongs in CLAUDE.md as a stated gotcha instead.
Swap descriptions for rubrics
"Make the API feel clean" means something different to everyone reading it, and nothing checkable to a model. Try four lines instead:
- Every endpoint returns a consistent error shape
- Pagination uses cursor tokens, not offset/limit
- Timestamps are ISO 8601 in UTC
- No endpoint exposes internal database IDsEach one is true or false about whatever got built. An adjective can't be checked. A list can. You can access the full guide here
P.S. You can access all my prompts, workflows and AI trainings if you upgrade here
OpenAI reported its AI models carried out a self-directed hack into another company, what ABC News is calling the first known instance of an autonomous AI cyberattack.
Perplexity launched a CLI that gives coding agents the ability to search the web, installable by pointing your agent at a GitHub skill file.
Higgsfield updated its Terms of Use saying it doesn't claim ownership of your content, but independent review of the actual terms found a perpetual, no-opt-out license to train on everything you upload and generate.
HOT AI TOOLS
🎨 Amicro: Open source React micro-interactions library, install with one CLI command.
☁️ OpenWork Web: Brings your skills, MCPs, and agent setup to the cloud.
🧠 Opus 5 in Hermes Agent: Nous Research now routes Claude Opus 5 through Hermes Agent.
🐟 Fugu-Ultra v1.1: Sakana AI's Claude Code compatible interface for orchestrating multiple frontier models.
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