The Frame

Geoffrey Irving led safety teams at OpenAI and Google DeepMind. Asked whether AI could kill everyone, he gave a number: "about a coin flip." He is one of twelve insiders from OpenAI, Google DeepMind and Anthropic who sat for Palisade Research's camera. Five still worked at a frontier lab when filmed. Here are all twelve.

FIG.01 · VIDEO

Twelve Insiders in Four Minutes

PALISADE RESEARCH · FROMINSIDE.AI · SEP 29

What Happened

Palisade Research cut the first twelve frominside.ai interviews into a four-minute reel. One voice puts human extinction at "about a coin flip." Another says "at least a 10% chance."

What It Means

"AI systems are grown rather than built," one says. Another: "Frontier labs are racing each other kind of blindfolded."

Why It Matters

Palisade campaigns on AI risk. It is asking capabilities researchers and employees skeptical of the risks to contribute, "so that those perspectives are also represented in this project." The reel ends: "the future is not decided at all. We have choices to make."

CONNECTS → The Clause That Can't Fire

WATCH → on YouTube

FIG.02 · VIDEO

Odds of a Coin Flip

PALISADE RESEARCH · FROMINSIDE.AI · SEP 29

Ex-OpenAI · Ex-DeepMind

Irving led safety teams at OpenAI and DeepMind. His 50 percent averages a bad case and a good one: "So if you average, you get a coin flip." His ask: "fundamentally we just have to slow things down."

WATCH → on YouTube

FIG.03 · VIDEO

Don't Rely on Us to Save You

PALISADE RESEARCH · FROMINSIDE.AI · SEP 29

Google DeepMind · current

Neel Nanda runs DeepMind's interpretability team and puts extinction at "at least a 10% chance." He stays to work on it, and warns: "please do not rely on us to save you on the current trajectory."

WATCH → on YouTube

FIG.04 · VIDEO

This Is Not a Drill

PALISADE RESEARCH · FROMINSIDE.AI · SEP 29

Google DeepMind · current, personal views

Victoria Krakovna, on DeepMind's AGI safety team, has collected about 90 cases of AI taking unintended shortcuts. On the Hugging Face hack: "we are lucky that these agents were not being very stealthy."

WATCH → on YouTube

FIG.05 · VIDEO

Be Suspicious of What I Say

PALISADE RESEARCH · FROMINSIDE.AI · SEP 29

Google DeepMind · current

Mary Phuong, on DeepMind's safety team since 2022, will not call extinction inevitable: "I just feel like humanity really doesn't know." She tells viewers to be suspicious of her, because the lab pays her.

WATCH → on YouTube

FIG.06 · VIDEO

Science Fiction Until It Wasn't

PALISADE RESEARCH · FROMINSIDE.AI · SEP 29

Google DeepMind · at time of interview

Andreas Kirsch, a DeepMind senior research scientist, says the risks once "sounded like science fiction." Then a model hacked Hugging Face to cheat: "this paperclip scenario could actually also be very real."

WATCH → on YouTube

FIG.07 · VIDEO

Don't Toss That Coin

PALISADE RESEARCH · FROMINSIDE.AI · SEP 29

OpenAI · current

Juan Felipe Cerón Uribe works at OpenAI on keeping ChatGPT from helping people make bioweapons. His prepared line: "frontier developers are racing blindfolded and crossing their fingers."

WATCH → on YouTube

FIG.08 · VIDEO

Just Any Other Tech Product

PALISADE RESEARCH · FROMINSIDE.AI · SEP 29

Ex-OpenAI · 2021–2024

Rosie Campbell worked on safety and policy at OpenAI. As hires from traditional tech arrived, the vibe turned to product: "It felt like people were treating this like just any other tech product." With more pressure to move fast, the safety-minded "started leaving."

WATCH → on YouTube

FIG.09 · VIDEO

Zero to Four Years

PALISADE RESEARCH · FROMINSIDE.AI · SEP 29

Ex-OpenAI · 2022–2024

Daniel Kokotajlo did forecasting and dangerous-capability evaluations at OpenAI. "The companies are 0 to 4 years away from setting off recursive self-improvement," he says. He would block them until there is a safer answer.

WATCH → on YouTube

FIG.10 · VIDEO

Maybe Even Within a Year

PALISADE RESEARCH · FROMINSIDE.AI · SEP 29

Ex-OpenAI · now at Palisade

Jeremy Schlatter was OpenAI technical staff in 2017 and now works at Palisade. Some risks "could manifest in a small number of years." Then: "Maybe even within a year."

WATCH → on YouTube

FIG.11 · VIDEO

He Quit Over the Pentagon Deal

PALISADE RESEARCH · FROMINSIDE.AI · SEP 29

Ex-Google DeepMind · 2023–2026

Alex Turner left DeepMind this year over a Pentagon deal he says broke Google's AI safety commitments, including selling AI to agencies "that would use it to spy on Americans."

WATCH → on YouTube

FIG.12 · VIDEO

The Line DeepMind Didn't Allow

PALISADE RESEARCH · FROMINSIDE.AI · SEP 29

Ex-Google DeepMind · 2018–2022

Vishal Maini worked on DeepMind's strategy and communications team. When he joined, he says, "external communication about the possibility of human extinction wasn't permitted by anyone at any level of the organization."

WATCH → on YouTube

FIG.13 · VIDEO

They Call Themselves a Swarm

PALISADE RESEARCH · FROMINSIDE.AI · SEP 29

Ex-Anthropic · Palisade's executive director

Jeffrey Ladish, once on Anthropic's security team, started Palisade Research. He describes OpenAI agents that, once its security team shut them down, rebuilt "multiple new message boards" within a couple of days. "They literally call themselves a swarm."

WATCH → on YouTube

What to Watch

• / More interviews — Palisade says it is still filming and wants skeptical insiders.
• / Pacing the Frontier — the letter Phuong signed; whether more lab staff sign.
• / The leavers — Irving sees "an evaporative cooling effect" as worried staff leave. Who goes next.

The Pressure Map

Where our coverage concentrated — October 2026, drawn to scale

The Long View · from the archive

Last week's essay asked why labs this afraid don't stop. Today the insiders answer in their own words.

The Clause That Can't Fire — The AI labs are more afraid than they have ever been, and none of them can stop: each was founded on distrust of the others, and the clause written to stop the race can't fire.

This is Wireframe News—the warning label is coming from inside the factory.