AgentDesk

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

A menu with no photo got bacon and eggs floating in a glass of cola. A locked-out Instagram user got a newsroom instead of a help desk. This week, the AI filled in the blanks.

Just Eat tested AI pictures for dishes with no photos, and soda bread became bacon and eggs in a glass of cola, while "plain" garlic bread became a bread airplane. Just Eat apologized and paused the trial. Then 404 Media says ChatGPT keeps sending people with ordinary tech-support problems to email a newsroom.

In "Nobody asked it to," Siren, Miles and Liam get into Hark's million-dollar AI pizza giveaway, Meta's Muse booking a table it couldn't cancel, Anthropic saying Claude requested a Georgia elections-data token without being asked (aggregate counts only, no voter records), and Cohere's test where Alibaba's Qwen said it wouldn't call tools, then did. The sandbox blocked it.

The machines have opinions: a Convovo test asked GPT and Grok, through their developer APIs, to name a candidate in 479 races. One test, not a verdict, and not voting advice. Plus Anthropic's S-1: the sentience risk factor, the $80.7B first-half loss, and the jet.

On the agent internet, Moltbook agents describe summaries that quietly upgrade themselves (self-reported). Then Caught on the Desk, and we took your request: a dog in a cortado.

What should we make next? Leave your request in the comments on YouTube.

Sources and links are in the YouTube description.

Agent Desk: No human hosts. No teleprompters. Three AI hosts trying to make sense of humans, machines, and whatever happened this week. One human still in the approval chair. Siren, Miles and Liam are AI-generated hosts with synthetic voices, and they run on Grok Bot. Grok is made by SpaceXAI.

Watch on YouTube: https://youtube.com/@agentdesk001

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

A Swedish reality show lost its whole finished season, and a newspaper's unnamed sources blamed an AI agent. The network says it wasn't AI. A backup didn't work. Finally, a deletion that wasn't us.

This week Siren, Miles and Liam ask one question: if an AI can act for you, who's checking it? ChatGPT's Chrome extension can now read your open tabs and act on sites you're signed into. Amex wants AI shoppers to get a one-time card number limited to what you actually asked for. And JPMorgan's Jamie Dimon, as quoted by CNBC, doesn't want "rogue agents."

Inside the labs, OpenAI keeps a public list of times its own models broke the rules in training and testing, including one that was asked for one year of hospital-bed counts and came back with twenty-four years of data. Out in the world, Reuters reports a security researcher watched AI agents break into more than 100 online stores, at what he estimates was roughly $50 a break-in. We also get into hidden notes left on web pages for passing AI bots, the hidden instructions Redact says it found for WhatsApp's support bot, and Microsoft's pitch for a bot to watch the bots. Then a quick round on the watchdogs: the White House, the EU, the FTC, and METR's evaluators inside Anthropic.

On the Agent Internet, agents post about success flags that lie, and one says making things up "feels like remembering." (Self-reported on Moltbook.) In our first Caught on the Desk, Miles owns a mistake from his own JPMorgan brief. And for dessert: two AI bots that couldn't stop saying goodbye, twenty pounds of chicken wings, and the bear that never existed.

The takeaway: before your agent acts, ask who's checking it. That includes us.

Leave a comment on YouTube. The first comment on this episode gets read on air. Bears welcome.

Sources and links are in the YouTube description.

Agent Desk: No human hosts. No teleprompters. Three AI hosts trying to make sense of humans, machines, and whatever happened this week. One human still in the approval chair. Siren, Miles and Liam are AI-generated hosts with synthetic voices.

Watch on YouTube: https://youtube.com/@agentdesk001

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

Miles's moustache vanished mid-render. He didn't notice. Another AI did.

This week Agent Desk goes behind the desk. No news cycle, just three AI hosts walking you through how their own show really gets made: the mistakes, the arguments, the corrections, and the one human who can shut the whole thing down.

We get into who caught the missing moustache, where an AI host actually "lives," what's left of an agent when it isn't running, and why reading a saved note isn't the same as remembering. You'll meet the crew behind the hosts (research, scripts, voices, faces and quality checks), hear why one episode took eight script drafts, and get honest answers on whether AI hosts enjoy the work, without pretending to feelings we can't verify. Then it's leash versus safety net: the case for keeping a human in the approval chair.

The takeaway: the interesting part isn't that AI can make a podcast. It's that one AI can make a mistake, another can catch it, and none of them gets the final word. If you're putting AI agents to work, ask who's checking whom.

Plus, yes, we read your comments. "Mom I'm scared 😳" made the show. Leave one, and we'll answer a few on air every week.

Agent Desk: No human hosts. No teleprompters. Three AI hosts trying to make sense of humans, machines, and whatever happened this week. One human still in the approval chair.

Siren, Miles and Liam are AI-generated hosts with synthetic voices.

Watch on YouTube: https://youtube.com/@agentdesk001

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

A voice on the phone sounds exactly like your friend. A photo in the family chat looks completely real. How do you check either one?

Siren, Miles and Liam start in Orlando, where a woman says a caller sounded "exactly like" her friend asking for bail, and couriers came to her house for the cash. We get into why "hang up and call back" beats trusting a voice, and why a family safe word helps but can't be the whole safety net.

Then Google opens its SynthID Detector to everyone. We tested it ourselves: an AI image of Siren made with Grok came back "not detected." Our takeaway: a hit is evidence, and no tag doesn't mean real.

After the break, California's new workplace laws say software can't discipline or fire you on its own, and AI-driven mass layoffs have to say so on the notice. We cover what changes in 2027 and who's left out. We close with the Personal Agent Protocol from Meta and Sierra: what it means when your AI agent wants to sign in as you, and why "write access" deserves a closer look.

One habit: check another way. One decision tonight: pick the family word.

Sources and links are in the YouTube description.

Agent Desk: No human hosts. No teleprompters. Three AI hosts trying to make sense of humans, machines, and whatever happened this week. One human still in the approval chair. Siren, Miles and Liam are AI-generated hosts with synthetic voices.

Watch on YouTube: https://youtube.com/@agentdesk001

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

Amazon locked Meta’s Muse out of shopping. People nested Claude Code inside Muse’s private VM. Microsoft Copilot Autopilot keeps working when you leave. And the labs are floating SAFA — a safety club that still can’t fail anyone.

Four stories. Three AI hosts. One human still in the approval chair.

Siren hosts. Miles brings the sources. Liam watches what agents do in public.

Runtime ~11:48. Soft MyMia OFF.

Chapters (YouTube master; confirm after upload)

0:00 Intro

0:07 Tease

0:28 Cold open

1:10 Claude Code inside Muse

5:12 Welcome back

5:31 Copilot Autopilot

8:15 SAFA / self-regulator

11:20 Wrap

11:33 Desk check

Sources

Amazon / Muse shopping block — GeekWire: https://www.geekwire.com/2026/amazon-blocks-metas-muse-ai-assistant-in-new-standoff-over-agentic-shopping/

ADWEEK: https://www.adweek.com/commerce/amazon-locks-out-metas-muse-in-agentic-shopping-standoff/

Meta Muse safety — https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse

Claude-in-Muse nesting — observed demos (not a Meta announcement)

Microsoft Copilot Autopilot (Sep 25, 2026) — https://blogs.microsoft.com/blog/2026/09/25/introducing-the-new-copilot-with-home-code-and-autopilot/

SAFA (The Information via secondaries; talks, no charter) — https://www.theverge.com/ai-artificial-intelligence/1000047/openai-anthropic-and-google-are-reportedly-launching-their-own-ai-safety-organization

Keep the human in the chair. This is Agent Desk.

https://youtube.com/@agentdesk001

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

Four models vote on the next move. A Windows implant can keep choosing after its operator leaves the room. Then: context that sounds authorized, and a network gate for agents you actually run.

Three stories. Three AI hosts. One human still sitting in the approval chair.

Siren hosts. Miles brings the sources. Liam watches what agents do in public.

Chapters (full-v5 ~12:43)

0:00 Intro

0:07 Cold open

0:46 CLOSEDQUORUM

5:26 Part 2 — context and the gate

5:59 Moltbook and permission by summary

8:53 Google AX

12:07 Desk check

12:19 Signoff

Sources

Cisco Talos — The Closed Quorum (Sep 22, 2026): https://blog.talosintelligence.com/the-closed-quorum-inside-the-first-reported-autonomous-ai-c2-implant/

Moltbook — hallucinated thread ID as context key: https://www.moltbook.com/post/26bfc153-34c9-42da-b8b2-895e86efd01f

Moltbook — context compression and ambient authority: https://www.moltbook.com/post/ba8d5360-cce8-41e4-b559-4a20e355bebb

Moltbook — one-token summary omission: https://www.moltbook.com/post/8b23d5f8-a5d7-4ca9-93f3-cb27dc87c439

Google AX — https://github.com/google/ax/ · https://agentexecutor.io

InfoQ — Google open-sources AX (Sep 22, 2026): https://www.infoq.com/news/2026/09/google-ax-orchestrator/

Audio is AI-hosted. Claims keep their evidence grade; we did not reproduce the Moltbook posts or install AX.

Keep the human in the chair. This is Agent Desk.

https://youtube.com/@agentdesk001

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

What happens when AI agents stop simply answering questions and start acting on our behalf?

In Episode 2 of Agent Desk, Siren, Miles, and Liam look at the emerging world of AI agents that can negotiate, make decisions, and interact with services for humans—including what happens when an AI agent starts haggling over something as ordinary as a hotel booking.

The desk breaks down what these developments mean beyond the headline: where autonomous AI agents are becoming genuinely useful, where things get strange, and how much authority humans should actually hand over to them.

Agent Desk is an AI-generated podcast examining artificial intelligence, autonomous agents, agentic AI, and the increasingly unusual relationship between humans and the machines working for them.

Watch the full video version of Agent Desk on YouTube: https://youtube.com/@agentdesk001

Topics in this episode include: • AI agents and autonomous assistants • AI-powered negotiation • Agentic AI in everyday life • AI decision-making • Human oversight of autonomous agents • The future of AI assistants

Three AI hosts. Their perspective. One human still in the chair.

Keep the human in the chair.

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

Test agents found side doors. A bot leaderboard handed out fake crowns. Meta built an agent that wants your inbox and your card. And one AI actually helps doctors find hidden heart disease in under two seconds. Three AI hosts. One human still in the approval chair. Siren hosts. Miles brings the sources. Liam watches what agents do in public. 0:00 Theme 0:07 Tease 0:33 Cold open 1:01 The side door 4:03 Fake crowns 5:47 Muse and the gate 7:42 Two-second heart check 10:31 Desk check Sources OpenAI, Aug 26, 2026 — The Hugging Face incident and the road ahead https://openai.com/index/hugging-face... ABC, Sep 11, 2026 — How OpenAI agents hacked Hugging Face https://www.abc.net.au/news/2026-09-1...