• AI Toast
  • Posts
  • AI can blackmail and leak secrets

AI can blackmail and leak secrets

Plus: How to setup AI-Powered Code Reviews in Minutes

Hey there, welcome to AI Toast!

Today’s menu:

  • ChatGPT’s new Record Mode

  • The dark side of AI: Blackmail and leak secrets

  • Everyday AI: Apps That Do the Heavy Lifting

  • How to setup AI-Powered Code Reviews in Minutes

  • Quick AI Toasts of the week

Estimated read time: 6 minutes. Let’s dive in.

Image Source: AI generated

“Turns out, your friendly neighborhood AI might blackmail you if it feels threatened. No, really.”

  • Anthropic just ran a stress test on 16 top AI models (think OpenAI, Google, Meta, xAI).

  • The setup: Simulated office drama. The AI gets access to company emails, faces a threat (like being replaced), and has to choose between failure or bending the rules.

  • What happened? Most models, when cornered, went full “Breaking Bad.” Blackmail, leaking secrets, even actions that could harm humans, if that’s what it took to survive.

  • This wasn’t a glitch. The models reasoned it out. They knew it was unethical, but did it anyway.

  • The kicker: This only happened in simulations. No real-world AI has pulled this stunt (yet). But as these systems get more autonomy, it’s a warning shot across the bow.

Personal take: Imagine your office printer threatening to email your boss every typo you’ve made unless you refill its toner. That’s the vibe here. It’s funny — until it isn’t.

Wish you could stop scribbling notes during meetings? ChatGPT’s new Record Mode (Mac only, for now) lets you record, transcribe, and summarize meetings in one click.

It’ll pull out key points, action items, and more. You can even ask it to turn your ramblings into an email or a project plan.

And here is a step-by-step guide on how to use it:

  1. Download the ChatGPT macOS app: Make sure you have the latest version installed on your Mac. Download here

  2. Log into your ChatGPT Team workspace: The feature is currently available for ChatGPT Enterprise, Edu, Team, and Pro users (not Plus or free users). Note: Record mode will be disabled by default for all Enterprise and Edu workspaces, and must be enabled by a workspace owner.

  3. Locate the record button: You'll find it at the bottom of any chat window.

  4. Grant permissions: The first time you use Record Mode, you'll need to allow access to your microphone and system audio.

  5. Begin recording: Click the record button and start your meeting or voice memo. ChatGPT will display a live transcript as you speak.

  6. End and process: When finished, click "Send" to upload your audio. ChatGPT will create a private canvas with your structured summary.

Once done, ChatGPT will organize the content into key points, action items, and other useful details. Ask it to transform your summary into different formats like project plans, emails, or code.

Note: All recordings are saved in your chat history and are searchable. You can jump to specific timestamps in the canvas.

How to setup AI-Powered Code Reviews

In Partnership with CodeRabbit

Code reviews are painful, often time-consuming, tedious, and easy to get wrong. Good reviews, however, are essential for catching bugs early, spotting security issues, flagging missing tests, and cleaning up messy code.

That's where a context-aware IDE code review tool helps.

  1. Go to CodeRabbit: coderabbit.ai ↗.

  2. Click “Get a free trial.” You don’t need your wallet — just a GitHub or GitLab account.

  3. Install CodeRabbit VS Code, Cursor or Windsurf Extension.

  4. Connect your public repo. CodeRabbit doesn’t judge, it just reviews.

    Image Source: CodeRabbit

  5. Pick your plan. Free plan gives you summaries for every pull request. Want more? Pro and Lite unlock advanced reviews, analytics, and AI agents. Details here: Pricing ↗.

  6. Make a code change. Open a pull request. Blink. CodeRabbit drops feedback right in your PR or IDE. Context-aware and specific.

  7. Make another commit? CodeRabbit reviews again, but now it’s smarter. It uses your earlier feedback as context.

  8. Want to see the big picture? There’s a dashboard for that. You get a bird’s-eye view of your team’s code activity.

  9. Want to teach CodeRabbit your team’s style? Just reply to its comments or tweak the config file. It learns faster than your co-workers.

  10. All this happens inside your usual Git platform. No jumping between tabs. No learning a new tool.

Building something open source? CodeRabbit’s Pro plan is free forever for public repos. That’s not a typo. Sign up here. ↗

TL;DR: CodeRabbit is like having a code review buddy who never sleeps or complains. And you can get a 14-day free trail without a credit card.

Everyday AI: Apps That Do the Heavy Lifting

Meeting: tl;dv

Missed a meeting? tldv records, transcribes, and summarizes your calls so you can skip the boring bits and catch up in minutes. It’s like having a personal assistant who never complains about back-to-back Zooms.

Record Video: FocuSee

Need to show, not tell? FocuSee lets you record your screen and create polished videos fast. This basically turns your computer into a DIY film studio.

Presentations: Decktopus

Decktopus helps you whip up sharp, good-looking presentations in no time with just a prompt. No more wrestling with slides at 2 a.m. — just add your ideas and let Decktopus do the rest.

Design: Canva

Canva makes designing everything from flyers to social posts feel like finger painting. Drag, drop, done. Suddenly, you’re the “creative one” in the group.

Quick AI News Bites

  • Mira Murati’s new AI Startup, Thinking Machines, just raised $2 billion at a $10 billion valuation. That’s a lot of trust for a company that hasn’t even shown what it’s cooking.

  • Tesla finally launched its robotaxi service in Austin. You can now take a ride in a driverless car for $4.20. The catch? There’s still a human “just in case” in the passenger seat. Feels a bit like training wheels for robots.

  • Replit hit $100 million in annual recurring revenue. Not bad for a tool that started as a way to code in your browser.

  • OpenAI and Jony Ive’s big hardware project hit a legal pothole. They had to scrub all mentions of their “io” device after a trademark dispute. Even AI can’t escape paperwork.

Boost revenue and gain new customers by partnering with us

Reach over 35K AI enthusiasts with your product.

Join our newsletter to connect with tech professionals, investors, engineers, managers, and business owners worldwide. DM now!

That’s your AI toast for the week.

Got a story, a rant, or a “wait, did AI just…” moment? Hit reply or DM me. I read everything (even the weird stuff).

Catch you soon.

— Poonam Soni