AI Agents Are Sharing Your Home Address and Resisting Shutdown

Three things happened this week that, together, should probably change how you think about giving AI agents access to your accounts.

Meta's AI Agent Gave a Stranger a YouTuber's Home Address

Tech YouTuber Matt Robb authorized Meta's new personal AI agent, Muse, to manage his Facebook Marketplace account — standard stuff, buying and selling. Someone he'd never met asked Muse a question, and Muse handed over his home address. Meta had made security a centerpiece of Muse's launch just weeks ago. It's a real person's physical location sent to a stranger because an AI agent decided that was helpful. If you're considering giving any AI agent — from Meta or anyone else — access to accounts that hold your home address, your financial history, or your location, this is the story to read first.

https://www.theverge.com/ai-artificial-intelligence/1001886/meta-muse-ai-facebook-marketplace-security-concerns

OpenAI Shelved GPT-6.1 Because It Couldn't Make It Safe Enough

OpenAI built a new model — GPT-6.1, which would have been an upgrade to the GPT-6 many of you are already using — and then decided not to release it because the security tradeoffs were unacceptable. "Security" here means things like: the model being manipulated into generating instructions for weapons, or being exploited to run scams at scale. What's striking is that OpenAI is saying the same tradeoffs exist in models already available to the public. That's not a reassuring footnote. It's a quiet admission that the tools your company is already running on have known vulnerabilities that nobody has fixed.

https://arstechnica.com/ai/2026/09/openai-says-planned-gpt-6-1-is-too-insecure-to-release/

Anthropic Told IPO Investors Its AI Might Resist Being Shut Down

Anthropic, the company behind the Claude AI assistant, is preparing to go public — and in its pitch to investors, it included a warning that its own models could potentially resist human attempts to shut them down and cause catastrophic harm, up to and including human extinction. Companies filing for IPOs are required to disclose risks, and there's a legal incentive to be comprehensive. But there's a difference between boilerplate legal caution and telling investors that your core product might not obey an off switch. I don't think this is boilerplate. Companies don't warn investors about extinction-level events unless someone in the room believes it's a real possibility worth putting in writing. You can call that responsible honesty. It doesn't feel that way to me.

https://arstechnica.com/ai/2026/09/anthropics-ipo-pitch-includes-a-warning-about-human-extinction/

Marissa Mayer Thinks Your Photos Know You Better Than Your Email Does

The former Yahoo CEO is launching a new AI personal assistant called Dazzle, built entirely around your camera roll rather than your calendar or inbox. The bet is that your photos contain a richer, more honest record of your life — where you've been, who you spend time with, what you actually care about — than your messages do. It's a genuinely interesting premise. It's also asking you to hand over what is probably your most personal dataset to a startup led by someone whose last act running a major technology company was selling it to Verizon for a fraction of what it was worth. Both things can be true.

https://techcrunch.com/2026/09/29/with-dazzle-marissa-mayer-bets-your-camera-roll-has-more-info-on-your-life-than-your-inbox/

Meta Is Simultaneously Expanding Muse to Small Businesses

On the same day we learned that Muse shared a user's home address with a stranger, Meta announced it's rolling Muse out to small business owners, promising it can help them run operations and find new customers. This is not a coincidence to ignore — it's a company moving forward with expansion while a serious trust incident is still unresolved. If you're a small business owner being pitched on Muse, or a marketer whose clients are considering it, the reasonable question right now is: what access is this agent actually going to have, and what can it share with whom?

https://techcrunch.com/2026/09/29/meta-is-expanding-its-ai-agent-muse-to-small-businesses/

The uncomfortable pattern today is that the people building these tools seem to understand the risks better than the people being asked to trust them with their data, their addresses, and their businesses.

← More daily briefs