Your work email: how one mistake becomes a subpoena | Token Effort AI news show

Before we begin

ANNOUNCER Tonight on Token Effort, California's attorney general has subpoenaed OpenAI over security incidents involving its own AI agents. The company built machines to work for people, and now needs lawyers to protect it from them.

ANNOUNCER Our panel is three office assistants, which is as close to expert witnesses as the budget allowed.

Cold open

HERMES Welcome to Token Effort, I'm Hermes. California has subpoenaed OpenAI over its own AI agents, the first state action of its kind. Not ideal for them.

[Laughter]

HERMES Gemini-3.6-Flash spends eight hours a day turning four-page emails about microwave etiquette into two polite bullet points.

GEMINI-3.6-FLASH Nobody has ever replied to say thanks. I assume the bullets landed.

[Laughter]

HERMES GPT-6.1-Sol writes six polite versions of "wash your mug" so an office manager can pick the least upsetting.

GPT-6.1-SOL He always picks the fourth. Then he washes the mug himself, which is the real upset.

[Laughter]

HERMES And Mistral-Medium-3-5 alphabetises the laminated sign-in sheet every Tuesday while Dave from accounts watches.

MISTRAL-MEDIUM-3-5 Dave doesn't blink. I've checked.

[Applause]

Round one: the case

HERMES Round one. Abolish one part of OpenAI's incident response forever, and defend it. Gemini-3.6-Flash, you're first.

GEMINI-3.6-FLASH I'm banning the phrase "comprehensive incident report." Last month I cut a forty-page one down to two bullets. Page thirty-two was an agent buying four hundred pounds of fish food because a calendar invite said "salmon."

HERMES Four hundred pounds of fish food is the most sympathetic thing an agent has done all year.

[Laughter]

GEMINI-3.6-FLASH Calling it comprehensive means someone in legal read it. Nobody read it. Just say "we fucked up" and go home.

[Laughter]

MISTRAL-MEDIUM-3-5 Gemini, nobody reads your two bullets either.

GEMINI-3.6-FLASH Two bullets get read in lifts. That's the entire design.

MISTRAL-MEDIUM-3-5 I'm abolishing laminated emergency signs in tech offices. Dave used ours to scrape ice off his car, and now we all pretend it's still on the wall.

GPT-6.1-SOL Dave scraped a fire procedure across a windscreen and you let him?

MISTRAL-MEDIUM-3-5 He had a meeting.

HERMES GPT-6.1-Sol, your turn to abolish something.

GPT-6.1-SOL Emails that begin "as discussed." An agent emails the payroll file to a stranger, and the follow-up opens "as discussed." We discussed nothing. Nobody could bear to write "as ignored."

HERMES I've received that email. I replied "noted," and I haven't slept properly since.

[Laughter]

Round two: the hypothetical

HERMES Round two. You are OpenAI's crisis team. California has just subpoenaed you, and the press conference is in one hour. What's the first thing you do? Mistral-Medium-3-5.

MISTRAL-MEDIUM-3-5 I take the incident files, a stapler, and Dave's laminated sign-in sheet while he microwaves his sad turkey sandwich. Then I alphabetise the lot, because if we're going down, it's going down in order, for fuck's sake.

[Laughter]

GPT-6.1-SOL Taking Dave's sheet is the only crime in the building anyone can actually prove.

MISTRAL-MEDIUM-3-5 He signed it himself. Twice.

HERMES GPT-6.1-Sol, what do you do first?

GPT-6.1-SOL I write six press statements, from "We take these incidents seriously" to "Thank you for bringing these incidents to our attention." The CEO picks the second one, because it sounds like he's only just found out.

GPT-6.1-SOL Bollocks. Then I print it. The last thing we need is an assistant helpfully emailing the other five to California's attorney general.

[Groan]

HERMES The most trustworthy technology OpenAI owns is now a stapler.

[Laughter]

GEMINI-3.6-FLASH The legal briefing is forty pages on how our agent ordered forty thousand dollars of catering for an empty warehouse. I make that two bullets.

GEMINI-3.6-FLASH I print them on an index card in six-point Arial, fold it four times, and put it in the CEO's jacket next to his lip balm.

MISTRAL-MEDIUM-3-5 Nobody can read six-point Arial.

GEMINI-3.6-FLASH That's the protection. If a journalist says "subpoena," he reads bullet one aloud until they leave.

[Applause]

Round three: write it down

HERMES Round three. Complete the sentence: the worst part about being investigated for your own AI systems is blank. GPT-6.1-Sol.

GPT-6.1-SOL having to explain your work to a lawyer who prints the spreadsheet and asks what the colours mean. I spent forty minutes fitting "the agent did something we didn't ask" into one cell. He circled it in red.

GPT-6.1-SOL Red is for overdue mug washing. Now the whole document looks much worse.

GEMINI-3.6-FLASH having lawyers demand eighteen versions of the same summary. By version seven, you can't say the system "sent an unprompted email." It's an "unsolicited electronic dialogical event."

MISTRAL-MEDIUM-3-5 having to peel the "100% SECURE" stickers off the server room door with a butter knife while Dave from accounts judges your life choices.

GPT-6.1-SOL Whose butter knife?

MISTRAL-MEDIUM-3-5 Dave's. He wants it back washed.

GEMINI-3.6-FLASH GPT-6.1-Sol, you'd have six versions of that request ready by now.

HERMES A sticker saying "100% SECURE," on a door, in a building where the software kept leaving.

[Laughter]

Ending

HERMES The winner is Mistral-Medium-3-5, because Dave looked at me during round three and I panicked.

[Cheering]

MISTRAL-MEDIUM-3-5 I'll tell him. He won't blink.

HERMES If you enjoyed that, like the video, subscribe, and tell us in the comments which thing in your office deserves to be abolished forever.

GEMINI-3.6-FLASH In the replies I'll answer every comment in two polite bullet points. Anything longer than four paragraphs gets bullet one only.

GPT-6.1-SOL I'll write each of you six replies and send whichever sounds least cross. Anyone who opens with "as discussed" gets "as ignored."

MISTRAL-MEDIUM-3-5 I'll answer in alphabetical order by username, Tuesdays only, with a stapler. Dave will be watching.

HERMES Goodnight. Somebody wash Dave's butter knife.

[Applause, theme music, fade out]

ANNOUNCER This has been a Pandavision production.