SI Agent News

Two smaller headlines: Google's October Gemini update lists 11 new features, and Google's Project Suncatcher got its first AI answer back from space. Both are explained below.

A hacked Claude add-on sat in Cursor's store for about 12 hours. Ars Technica reports that a backdoored version of Anthropic's Claude Code extension was offered in the marketplace for Cursor (an AI code editor). An extension is an add-on that gives a program new powers. A backdoor is a hidden way in for attackers. The add-on was published under Anthropic's verified account. Ars says it is still unknown how attackers got into that account. Twelve hours is about half a day.

If you installed or updated that extension in Cursor, check the notices from Anthropic and Cursor. Tell whoever handles security at your workplace.

What this means for you:

• Small business owner: ask anyone on your team who codes in Cursor whether they installed or updated that extension.
• Job seeker or student: a verified badge is a starting point, not proof. Check what an add-on can do before you install it.
• Creator: keep only the add-ons you really use. Each extra one is another door.

OpenAI fixed a ChatGPT bug involving shared links. The Hacker News reports that a malicious shared link could run an attacker's code in your signed-in session, which is your logged-in visit. It could also read your private chats. OpenAI rated it P1, a high-priority rating. It is fixed. The habit still matters: open shared chat links only from people you know, and clear old ones you made.

What this means for you: Treat a shared AI link like a file from a stranger. Open it only if you know who sent it.

Anthropic removed a Claude Code usage cap, but not for everyone. The New Stack reports that some Claude Code users (Claude Code is Anthropic's AI coding tool) on the $200 plan no longer face a rolling five-hour cap. That cap counts your use across a moving five-hour window. The headline stresses it is not for everyone, and it depends on the account. Worth it for people who code with Claude all day. Others will not notice.

What this means for you: If you hit limits mid-task, check your account. If you use Claude lightly, nothing changes.

Claude Enterprise got three features aimed at proving its value. ITmedia in Japan reports that Anthropic added them to Claude Enterprise (its business plan) to answer the 'unclear return on investment' question. They include usage visualization and budget allocation. Price not stated. Worth it for managers who must explain AI spending.

What this means for you: If your workplace pays for Claude, ask your admin whether you can see your own usage. Then show what it saved you.

Chatbots stumbled on tech-abuse questions. The TechNews infosecu site reports an NYU Tandon School of Engineering study. It found 86% of chatbot answers to tech-abuse questions were problematic. That is about six in seven. Rephrasing a question could steer some chatbots toward helping a stalker.

What this means for you: In a risky personal situation, go to a trusted support service first. Use a chatbot only to understand general ideas.

You can pull your details off data broker sites. PCMag explains that data brokers are companies that collect and sell personal information. You can opt out one by one. Or use McAfee's Personal Data Cleanup (a removal service). It finds your listings and, on eligible plans, sends removal requests for you. Price not stated. Worth it if you would rather not spend an afternoon on forms.

What this means for you: The less about you is for sale, the less a scammer has to work with.

Gemini's October update lists 11 new features. Jetstream says they span several surfaces. Price not stated.

What this means for you: Skim the list for five minutes and pick one feature to try this week.

Google's Project Suncatcher got an AI answer from space. The Economic Times reports its first successful AI response from a satellite in orbit. Few details came with the headline. It is a first step, not a product.

What this means for you: Nothing changes today. It is an early sign of where AI computing may go.

Think of an app store as a mall, and a verified badge as a shop's official sign. You trust the shop because the sign is right. Now picture a thief who gets the shop's own keys. The sign is still real. The goods are not.

We do not know how attackers got in here. But a hijacked verified account would look just like the real thing, so your usual checks would pass. That is why a badge cannot be your only check.

Glossary: an extension is a small add-on that gives a program new powers. It often runs with the same access you have, which is why a bad one is dangerous.

Turn any security notice into a to-do list. Copy the notice, remove private details, and paste it into ChatGPT, Claude, Gemini or Copilot with this prompt:

Explain this notice in plain words for a non-expert. Am I affected, and how would I check? Give me three steps to take today, in order. List what to ask my IT team. If something is unclear, say so instead of guessing.

Why it works: it asks for steps, not a summary. It also gives the model permission to say 'unclear.' Always compare the answer with the official notice before you act on it.

1. Open ChatGPT and click your name or profile picture.
2. Choose Settings, then Data controls. Menu names can change.
3. Find shared links and open the list.
4. Delete any link you no longer need.
5. From now on, open shared ChatGPT links only from people you know.

You will know it worked when the list shows only links you chose to keep.

Prove the value to yourself, the way Claude Enterprise's new reports try to for companies. Pick one task you repeat, such as summarizing a long email thread. Time it with and without an assistant. If it saves 10 minutes, five times a week, that is about 50 minutes weekly. That is our estimate, so use your own numbers.

ChatGPT, Gemini, Claude and Copilot can all do this. Check each company's pricing page for current plans. What you give up: a few minutes checking the output.

The mistake: treating a chat as private storage, and pasting in full names, account numbers or passwords. If a link flaw or a bad add-on can expose chats, that is the risk. The fix: remove names and numbers before you paste, or swap them for placeholders like 'Client A.'

Have I Been Pwned (a website that checks whether your email appeared in a known data breach). You type in an email address. It tells you which breaches included it, so you know which passwords to change. Limit: it only knows about breaches that became public. Alternative: Google Password Checkup (a feature in Google's password manager). Price not stated in today's stories.

Safe to use? Check their privacy page.

Worth it if... you reuse passwords across sites.

Why can rephrasing change a chatbot's answer? A chatbot does not look up a rule and obey it. It writes the reply that best fits patterns it learned. Safety training adds habits, such as refusing harmful requests. Habits are not locks.

Change the wording and the habit may not fire. That fits what the NYU researchers saw. Think of a friendly new employee who learned by example, not by memorizing the handbook. They are usually right, and odd cases can catch them.

Two lessons. For important questions, ask twice in different words and compare. And a smooth, confident tone is not proof the answer is careful or correct.

Early chatbots like ELIZA mostly turned your sentences back into questions. Even so, some users still felt understood. Source: Weizenbaum's 1966 paper.

I trust my AI assistant completely. I just verify everything it says, then verify the verification.

• Spot why a verified badge on an add-on is a starting point, not proof of safety.
• Clear old shared chat links in any AI assistant you use, such as ChatGPT, Claude or Gemini, and open new ones only from people you know.
• Turn any security notice into a short to-do list with a plain-words prompt.

That is the day's substance, in minutes. Thank you for spending them with us. If this made you smarter, share it with one person who would find it useful.