OpenAI's new agents, dots, are fuzzy balls with eyes that you name and dress. They also sign into websites with your saved passwords and read the email you connect. I went through OpenAI's pages to see what the face is covering. What a dot reads can train the model unless you find the switch, and websites are told nothing. I would not give one my logins or my card.

Chapters

• 00:00 Sam Altman's pitch, and what a dot is
• 01:42 Do you need one?
• 02:36 Where the face came from
• 04:33 Launch day, and why now
• 07:05 Under the costume: your data
• 08:37 A model that looks for another way
• 10:06 What the cute is for
• 11:09 What websites are told
• 12:39 Through the web, or over it
• 14:01 The verdict

Key Takeaways

1. The face is how it gets sold. OpenAI had characters on the side in April. Meta made the character the product, and three weeks later OpenAI did too.
2. What a dot reads can train the model by default. The setting is "Improve the model for everyone", and it covers the apps you connect, email included.
3. OpenAI publishes no way for a website to recognise a dot. Its instructions tell the user to try their computer when a website blocks one.

What to Do

• In ChatGPT, open Settings, then Data controls, and uncheck "Improve the model for everyone".
• Before you give any agent your logins or a card, ask what it does that you needed.
• If you run a website, send it to me at nohacks.co/ai-accuracy. I'll put one question your customers ask to the AI assistants and send you what they said. It's free.
• The newsletter is at nohacks.co/subscribe.

Sources & Links

From OpenAI

• Introducing dots
• Dots privacy FAQ
• Dots, computers and apps
• OpenAI's crawlers
• GPT-6 Astra test results

Mentioned in the episode

• Thomas Germain, BBC, on Meta's Muse
• The Tech Report, with Ed Zitron
• Better Offline
• FTC on Joe Camel, 1997

No Hacks runs no sponsorships and is funded by advisory and audit work.

If your website needs to work for machines as well as people, start with a fixed-scope Machine-First Architecture audit: https://nohacks.co/audit

Mark Zuckerberg says Meta's AI agent, Muse, needs to be discreet. The same month, some of the calls it made to businesses were placed by people in a call center. I went through what he told Joanna Stern and Alex Heath about Muse's privacy and security and sorted every claim by tense. What exists today, only Meta has checked. The protection from Meta itself is promised for later this year. My verdict: stay away from this. After the sign-off, the show moves to Saturday mornings.

Chapters

• 00:00 - Discreet AI, and calls made by people
• 00:59 - What Muse is, and the agentic web
• 02:33 - Joanna Stern's question
• 03:57 - Present tense: the virtual machine and the ads
• 05:45 - Passwords, Sentinel, Amazon and the Mac app
• 07:05 - Future tense: the locked version and discretion
• 10:10 - Who is asking: an ad company with a privacy record
• 11:40 - From the browser wars to the agent wars
• 14:42 - Two documents, 1993 and 2026
• 16:00 - After the episode: moving to Saturdays

Key Numbers

• 500,000+ people tried Muse in its first week, as reported by The Information
• 97.6% of Meta's 2025 revenue came from ads
• $5 billion: the FTC's 2019 privacy penalty on Facebook
• 87 million people's data reached Cambridge Analytica
• 1993: CERN put the web's code in the public domain
• "Later this year": when Meta says the version it cannot see inside arrives

Key Takeaways

1. Listen for the tense. What Muse does today is described only by Meta, and nobody outside Meta has published a check of it. The version Meta itself cannot see inside is promised for later.
2. Meta's plan for Muse is a small cut of every transaction. Zuckerberg said the cut will come from the businesses. If you run a store, that is you.
3. The early web worked because it was given away and ran on your own computer. This time the companies want to be in the middle from day one. It is the browser wars again, fought over the agent.

What to Do

• Before you connect your email, passwords or card to any agent, ask how it works today, in the present tense.
• If you run a store on Shopify, check Sales channels, then Agentic. Meta's agent and Google's AI may be on by default.
• Watch the full Joanna Stern interview.
• The No Hacks newsletter is at nohacks.co/subscribe.

Sources & Links

The interviews

• Joanna Stern, New Things with Joanna Stern
• Alex Heath, Sources podcast

Meta's own words

• Introducing Muse
• Meta's 2025 annual report

The record

• FTC, $5 billion penalty
• CERN, 30 years of a free and open web

No Hacks runs no sponsorships and is funded by advisory and audit work.

If your website needs to work for machines as well as people, start with a fixed-scope Machine-First Architecture audit: https://nohacks.co/audit

A hundred and seventy-two companies have put their name to A2A, the standard for AI agents at different companies to find each other and talk. I took the web address the project lists next to each name and asked 135 of them the question the standard is built around. Four answered. The reason is not technical: MCP needs one person to say yes, A2A needs two organisations.

Chapters

• 00:00 - AWS, Microsoft, Salesforce, SAP, and four answers
• 00:54 - What I was actually sold, and what I went looking for
• 01:31 - What counts as another party, and what does not
• 03:08 - Single-player, and why the reason is not technical
• 04:07 - MCP, November 2024: one person has to agree
• 06:56 - A2A, April 2025: two organisations have to agree
• 11:04 - What the agent card is, and why not having one is fine
• 12:56 - Who is actually at your server
• 15:35 - Forget the logos, watch your own logs
• 16:24 - After the episode: the race to the bottom

Key Numbers

• 172 companies listed on the A2A partners page
• 135 of them checked, both well-known paths, 270 requests
• 4 served an agent card. 131 served nothing
• 1 served it at the path the specification registers
• 14 returned HTTP 200 at that path. 12 were the homepage
• MCP: 1 person has to agree. A2A: 2 organisations

Key Takeaways

1. The visitor at your server is one person's assistant, not another company's agent. It fetches one thing, cannot ask a follow-up, and has no patience for a slow page. That is a smaller job than the one being described to you, and a different one.
2. A protocol that pays off alone spreads. One that needs a counterparty waits. MCP works the afternoon you wire it up. You can implement A2A perfectly and get nothing until somebody you do not control has done the same work.
3. Not having an agent card is not a failure. It is not a test, nothing scores you on it, and if no agent runs on your server there is nothing to publish.

What to Do

• Watch your own logs rather than the partner lists. The day something arrives that was clearly sent by another company's system, not a person, is the day this changed.
• Treat a 200 as nothing. Twelve of the fourteen I found were a homepage answering to any address.
• Answer clearly and quickly, and be honest about what you do not have. Most websites say yes to everything.
• The No Hacks newsletter is at nohacks.co/subscribe.

Sources & Links

The standard

• A2A specification, section 8.2 discovery
• A2A partners list
• IANA well-known URI registry

The other protocol

• Model Context Protocol specification
• Announcing A2A, Google, April 2025

No Hacks runs no sponsorships and is funded by advisory and audit work.

If your website needs to work for machines as well as people, start with a fixed-scope Machine-First Architecture audit: https://nohacks.co/audit