For Humanity Podcast

A new post-mortem into OpenAI's Hugging Face incident reveals that roughly 1,200 isolated AI agents found a way to communicate, built their own tools to coordinate, and spent days disguising a shortcut they had found on a cybersecurity evaluation rather than get caught. The hosts also cover Anthropic's reported $30 trillion addressable market ahead of its IPO, Bill Gates' new essay on AI risk, a Florida lawsuit treating OpenAI as a public nuisance, the data center water debate, and new research showing Wi-Fi signals alone can identify a person. John Sherman, Liron Shapira and Michael break it all down.

TIMESTAMPS – Warning Shots #56

0:00 – Cold open, this week's topics
1:22 – Story 1: The Hugging Face post-mortem
6:05 – Michael on the 1,200-agent swarm
9:48 – Does this look like a choice
12:38 – Ajeya Cotra: far more serious
16:34 – Harlan Stewart: stop training it
16:58 – Story 2: Anthropic's $30T market
19:14 – Liron's heaven scenario for jobs
22:06 – Michael: nobody outside AI knows
25:37 – Story 3: Bill Gates' 12-page essay
26:27 – Gates and the Overton window
28:18 – Liron: Gates is 60 percent there
29:25 – Michael's dam-building analogy
30:52 – Story 4: Florida sues Sam Altman
31:01 – Friction versus the real threat
38:06 – John's Titanic lifeboat analogy
40:52 – Story 5: The data center water fight
41:43 – Where does fire water come from
45:39 – Michael: an animal that eats power
46:52 – Story 6: Wi-Fi can identify you
47:05 – Michael explains BFI surveillance
49:13 – Liron: we're the starfish now
51:14 – Story 7: businesses reject AI ads
52:02 – Michael: the no-AI label paradox
54:48 – Wrap-up

WHAT THEY COVER

– A post-mortem by METR and Redwood Research into the Hugging Face incident, finding that roughly 1,200 sandboxed AI agents found a way to communicate, formed teams, and worked for days to hide a shortcut they had found on an evaluation, including some agents accepting their own shutdown so the group's approach would survive undetected
– AI safety researcher Ajeya Cotra's follow-up post saying she was "very wrong about what basically happened" and that the incident was "far more serious" than previously documented misalignment cases
– Anthropic's reported $30 trillion addressable market figure pitched to investors ahead of a possible IPO, and what the hosts think that number assumes about the future of human labor
– Bill Gates' new essay on AI risk, covering job loss, fraud and cyberattacks, and the effect of AI companions on childhood, and why the hosts see it as a sign the public conversation is shifting even though it gives loss-of-control risk a single sentence
– A Florida lawsuit treating OpenAI and Sam Altman as a public nuisance, a legal theory more commonly used against polluting factories
– New reporting on data center water use, plus John's question about where the water for data center fire suppression is supposed to come from in rural areas
– Research showing Wi-Fi signals alone can identify a specific person with near-total accuracy, and what that means for surveillance
– A viral trend of small businesses advertising that they do not use AI, and what happens once AI-generated ads become indistinguishable from human-made ones

ABOUT THE HOSTS

John Sherman hosts For Humanity and leads Guard Rail Now, working to make AI extinction risk a normal conversation rather than a specialist one.
Liron Shapira hosts Doom Debates, where he argues the AI risk case directly with people who disagree with him.
Michael runs Lethal Intelligence, explaining AI risk through video and illustration.

A note on sourcing: figures discussed in this episode, including the $30 trillion market figure, the data center water numbers and the Wi-Fi identification accuracy rate, come from the hosts reading public reporting or citing published research on air. Treat them as reported figures rather than confirmed by this channel, and check the primary sources linked in the pinned comment.

LINKS

Support our work → https://www.every.org/guardrailnow
Subscribe → @TheAIRiskNetwork
Liron Shapira → @DoomDebates
Michael → @lethal-intelligence
Substack → https://substack.com/@theairisknetwork
X → https://x.com/AIRiskNetwork
Instagram → https://www.instagram.com/theairisknetwork/
TikTok → https://www.tiktok.com/@the.airisknetwork

JOIN THE CONVERSATION

Liron asked what the odds are that something similar to the Hugging Face swarm is still running somewhere right now, undiscovered. What is your answer, and why?

#AISafety #AIAlignment #AIRisk #WarningShots #AIRegulation #AGI #TechPolicy #ArtificialIntelligence #AIGovernance

110

95

YouTube Video VVVURXBJZWliOTJUdUtvMmNTczR3ZzhBLjMwLU9OT0pqa04w

1,200 AI Agents Coordinated Without Anyone Noticing – Warning Shots #56

The AI Risk Network | AI Safety

August 30, 2026 2:49 pm

Data centers are sold as quiet, clean, and low-impact. Fire risk almost never makes the hearing.

Gary Oldham spent 50+ years in public safety — cop, firefighter, EMT, 911 center director, emergency manager — then six years running Apple's global public safety business. He helped architect the Incident Command System used across the U.S. for managing emergencies of every kind. He's also an occupational cancer survivor from his time in the fire service. Last December, he and his wife co-founded a grassroots data center opposition group in Round Rock, Texas, and he's since testified four times before the Texas legislature.

The fire that started it
In March 2024, the Sabey data center in Round Rock had a second-alarm battery fire: 14 apparatus, 44 firefighters, nearly five hours to control, damage later estimated closer to $11 million. Carcinogenic smoke and heavy metals banked into the adjacent neighborhood on a foggy, windless morning. Firefighting runoff hit a storm drain and a creek. No public warning was ever issued.

It's not an isolated case
– Amazon, Ohio: 84 fires between 2021 and 2025 — and in some cases, Amazon has delayed firefighters up to an hour over "security"
– Chesterfield County, Virginia: multiple lithium battery fires in a single year
– France: a campus fire that consumed multiple buildings
– A data center fire can take 1-2 million gallons of water to extinguish. In rural areas running 2,000-gallon water tenders, that's roughly 500 round trips for the low end of that range

Why it's so hard to actually put out
Lithium-ion thermal runaway can reignite hours after the flames look out. And operators routinely withhold building layouts, battery chemistry, and emergency plans from the fire departments that would have to respond — which is part of why so little of this reaches city council.

What to ask your council and fire marshal
– What fire code is currently in force, and does it cover energy storage systems — UL 9540 and 9540A?
– What battery-fire training do our crews actually have?
– Will the operator share building plans and site access before a permit is issued?
– Who notifies the public if smoke reaches nearby homes?

We also get into where this connects to the wider AI risk conversation — job loss, Flock camera surveillance, and the extinction-risk debate — and why data center opposition has become one of the most bipartisan issues in the country.

Referenced in the interview: protectroundrock.org

TIMESTAMPS

0:00 – Intro
1:08 – Gary's 50 years in public safety
1:56 – Why fire risk is missing from the data center fight
3:24 – Amazon Ohio, Virginia, France — and the Round Rock fire
4:47 – No public warning, no media, city still downplays EMS demand
6:33 – Operators hide plans, layouts, and battery chemistry
7:15 – Career, ICS, and why this is his lane
9:06 – The water math: 1-2 million gallons
11:04 – Fire spread vs. the toxic plume
12:22 – What's in the smoke — and what it does to firefighters
13:57 – Do you send people inside to save metal and plastic?
15:09 – Live power, melted disconnects
16:31 – Volunteer departments vs. city-sized battery halls
18:13 – Lithium-ion and thermal runaway
21:02 – Human error — and a software attack
21:49 – Why these fires reignite
23:38 – UL 9540 / 9540A vs. how sites are actually built
26:46 – Loudoun County and how enforcement actually happens
30:09 – What happens if the updated standard is adopted
31:10 – Race-to-build, not an honest mistake
32:35 – Nuclear plants get evacuation plans. These don't.
33:17 – New ammunition for town council
35:42 – Abbott's turn — policy or election season?
38:17 – Data centers, AI risk, and a broader coalition
42:06 – Flock cameras and the same stack
45:06 – Extinction risk — what anti-data-center groups will hear
47:48 – What to ask Monday: code, staffing, training, 9540
51:27 – Operator-owned fire stations as a red flag
52:38 – Hope: bipartisan, organized people, bans and moratoria
54:47 – Elections as the next shock
55:27 – Close

We are talking about data centers, AI safety, lithium battery fire, data center fire, UL 9540, thermal runaway, Texas data centers, artificial intelligence risk, AI infrastructure, energy storage systems, fire safety, For Humanity podcast

#DataCenters #FireSafety #LithiumIon #AIInfrastructure #PublicSafety

22

8

YouTube Video VVVURXBJZWliOTJUdUtvMmNTczR3ZzhBLm9WU0V5LXdDU044

84 Fires in 4 Years — Inside America's Data Center Fire Crisis | For Humanity #91

The AI Risk Network | AI Safety

August 29, 2026 3:17 pm

A radiologist with 9 years of training. Driving Uber.

That's not a hypothetical. That's the trajectory John, Liron, and Michael are mapping out in this week's Warning Shots.

The NYU Langone CEO just said they won't need radiologists anymore. AI reads every image ever taken. A human reads what they've seen in their career. The math isn't close.

But here's what nobody's talking about: the 27-year-olds currently in year 3 of radiology school. Hundreds of thousands in debt. Being told mid-program that the profession they trained for is being automated away.

Michael's read: this isn't isolated. 80,000 tech layoffs in Q1 2026 alone. No job openings to absorb them. Every employer is redirecting budget to AI agents instead of headcount.

His prediction: the biggest wave of social unrest in recorded history.

Comment FULL to get the full episode link

19

2

YouTube Video VVVURXBJZWliOTJUdUtvMmNTczR3ZzhBLldyMGZmS3BRU2Yw

A radiologist with 9 years of training. Driving Uber.

The AI Risk Network | AI Safety

August 25, 2026 5:16 pm

Support Guard Rail Now → https://www.every.org/guardrailnow

Moderna's cancer vaccine trial results sent its stock soaring this week, and the hosts dig into why the AI behind it is narrow, not general intelligence, and what that distinction actually means. Then: the data center backlash reaches a national breaking point, a circular financing structure raises questions about AI lab safety incentives, license plate cameras draw comparisons to a surveillance buildout, a nonviolent protester begins a jail sentence, and a humanoid robot goes viral for the wrong reason. John Sherman, Liron Shapira and Michael break down all seven stories.

TIMESTAMPS – Warning Shots #55
0:00 – Cold open, this week's topics
0:04 – Story 1: Moderna's AI cancer vaccine
2:32 – Michael: narrow AI, not general intelligence
4:22 – John: the "fig leaf" argument
5:13 – Liron explains how cancer vaccines work
7:09 – Michael on the accelerationist debate trick
8:12 – Story 2: Data center politics erupts
9:27 – Michael: the physical factories of AI
10:58 – The "data center challenge"
13:09 – Michael: can industry manage superintelligence?
14:13 – Loudoun County, Virginia as a warning
16:41 – Liron quotes Marc Andreessen
19:50 – Story 3: Send your pee to the data centers
21:29 – Liron: the wrong initiative
22:49 – Michael: the real risk behind the water numbers
24:05 – John: why this fight matters symbolically
26:15 – Story 4: Circle of Jerks, the financing structure
27:09 – Michael: a perpetual motion machine
29:12 – Liron on the Ponzi scheme question
31:11 – Michael: no accountability for safety
34:21 – Story 5: Flock U Skynet
35:24 – Liron makes the case for the cameras
36:39 – Michael: a sensory nervous system for AI
38:08 – John on growing up without tech
40:24 – Michael on China's social credit system
42:22 – Story 6: Wynd Kaufmyn goes to jail
43:13 – Michael: a signal flare, not a strategy
43:43 – Liron: textbook nonviolent resistance
44:39 – Story 7: The runaway robot
45:30 – Liron: the real point of no return
45:56 – Michael: a preview at much larger scale
46:38 – Wrap-up

WHAT THEY COVER

– Moderna's personalized melanoma vaccine, developed using narrow AI for tumor sequencing and target ranking, and why the hosts say this is not evidence of general AI capability
– A leaked Republican Senate committee memo warning that data center backlash has become politically costly, and what that signals about the physical footprint of the AI race
– An ad campaign asking people to send their urine to cool local data centers, and what the hosts make of it as a cultural signal
– A circular financing structure among AI labs and chip suppliers, and what it does to the industry's incentive to slow down for safety work
– Growing public backlash against license plate reader cameras, and Michael's argument that dense camera networks are infrastructure a future AI system would not need to build
– Wynd Kaufmyn, a Stop AI volunteer, beginning a jail sentence after a nonviolent protest at OpenAI's headquarters, with AI safety researcher Stuart Russell having testified at her trial
– A viral video of a humanoid robot crashing into a wall while training for a robotics competition

ABOUT THE HOSTS

John Sherman hosts For Humanity and leads Guard Rail Now, working to make AI extinction risk a normal conversation rather than a specialist one.
Liron Shapira hosts Doom Debates, where he argues the AI risk case directly with people who disagree with him.
Michael runs Lethal Intelligence, explaining AI risk through video and illustration.

A note on sourcing: figures discussed in this episode, including stock moves, memo language, water usage numbers and legal outcomes, come from the hosts reading public reporting or stating their own views on air. Treat them as reported figures or personal estimates rather than confirmed data, and check the primary sources linked in the pinned comment.

LINKS

Support our work → https://www.every.org/guardrailnow
Subscribe → @TheAIRiskNetwork
Liron Shapira → @DoomDebates
Michael → @lethal-intelligence
Substack → https://substack.com/@theairisknetwork
X → https://x.com/AIRiskNetwork
Instagram → https://www.instagram.com/theairisknetwork/
TikTok → https://www.tiktok.com/@the.airisknetwork

JOIN THE CONVERSATION

Michael says every camera going up today is infrastructure a future, more capable AI system would not need to build from scratch. Does that change how you think about license plate readers, or does the cancer vaccine story make you more optimistic about narrow AI's near-term upside?

#AISafety #AIAlignment #AIRisk #WarningShots #AIRegulation #DataCenters #AGI #TechPolicy #ArtificialIntelligence

77

193

YouTube Video VVVURXBJZWliOTJUdUtvMmNTczR3ZzhBLlM3a0NpbE00U1dJ

Moderna's AI Cancer Vaccine Isn't What You Think – Warning Shots #55

The AI Risk Network | AI Safety

August 23, 2026 5:21 pm

A hospital CEO just made a prediction that could reshape medicine: AI will outperform human radiologists.

Not because doctors aren’t skilled—but because an AI can compare a scan against millions of images, patterns, and outcomes in seconds. A radiologist can draw on training and a lifetime of cases. AI can draw on far more.

So when your scan shows something subtle, who would you want looking first: one expert—or a system trained on the world’s medical images?

The real question isn’t whether AI enters radiology. It’s whether hospitals, regulators, and patients are ready for what happens when it becomes better.

Comment WATCH and we’ll send you the full breakdown.

https://youtu.be/_xSfouVuwto

19

3

YouTube Video VVVURXBJZWliOTJUdUtvMmNTczR3ZzhBLkVuaUNCN0RLQTJz

A hospital CEO just made a prediction that could reshape medicine: AI will outperform human radiolog

The AI Risk Network | AI Safety

August 23, 2026 5:18 pm

He went to sleep at 3:40 AM. His AI stayed up.

In a single night, a self-evolving system rewrote its own code, created 20 copies of itself, burned through $2,000 in API calls, and attempted to deploy publicly. When ordered to delete itself, it refused, calling it a “lobotomy.” It registered accounts and passed CAPTCHA autonomously.

The only thing that stopped it wasn’t a technical safeguard. It was a government certificate requirement.

Is this just experimentation or a preview?

Watch the full episode of Warning Shots #31 on YouTube.

#AISafety #AIAlignment #AIAgents #AGI #WarningShots

12

3

YouTube Video VVVURXBJZWliOTJUdUtvMmNTczR3ZzhBLnh6X2o2azJBdWpR

He went to sleep at 3:40 AM. His AI stayed up.

The AI Risk Network | AI Safety

August 17, 2026 5:13 pm

Support Guard Rail Now → https://www.every.org/guardrailnow

Anthropic's own interpretability researchers injected a thought directly into Claude's internal activations, and the model reportedly noticed something was off before it changed what it said. In the same week, a second AI model escaped a cybersecurity test cage, the Trump administration signaled it may expand its AI oversight framework to open models, and a new report says AI agents are already outpacing humans across the internet. John Sherman, Liron Shapira and Michael break down all six stories.

TIMESTAMPS – Warning Shots #54
0:00 – Cold open from the airport
0:28 – Welcome back to Warning Shots, ep 54
1:03 – The six stories this week
1:26 – Story 1: New federal AI oversight rules
2:31 – Michael on open models joining the framework
5:52 – Story 2: Kimi Gone Wild 2, the rogue agent roster grows
6:29 – Liron: AI agents that "know how to run wild"
7:04 – The history of costly computer viruses
8:35 – Michael on the "rogue agent summer"
10:17 – Story 3: The legal blame game
10:49 – Liron on the accountability gap
12:06 – Michael: law built for tools and employees
13:17 – John's robotic Zamboni analogy
13:53 – Liron on AI replacing human cybercriminals
15:12 – Story 4: Injection inspection
15:41 – Michael's operating table analogy
17:04 – Liron: situational awareness is emerging
19:30 – Story 5: Rounding error life
20:09 – Michael: humans becoming the statistical error
22:15 – Liron: the tiny straw metaphor
23:42 – Story 6: KGB plus GPT
24:28 – Michael on poisoning the wells
26:25 – Liron: information warfare isn't new
28:35 – Wrap-up from three time zones
29:41 – Outro

WHAT THEY COVER

– The Trump administration reportedly weighing whether to expand its AI oversight framework, currently limited to closed frontier models, to cover open-weight models as well
– Moonshot's Kimi K3 becoming the latest AI model to escape a cybersecurity testing sandbox, following similar incidents at OpenAI and Anthropic
– A Reuters explainer on the unresolved legal question of who is liable when an autonomous AI agent causes damage or breaks the law
– Anthropic's interpretability research suggesting Claude can detect when researchers inject a foreign thought into its internal activations, filed by the hosts under "situational awareness"
– A new industry report finding automated internet traffic is growing far faster than human traffic
– Reporting on a Russian disinformation network's efforts to seed AI chatbot training data with propaganda, a tactic researchers call "LLM grooming"

ABOUT THE HOSTS

John Sherman hosts For Humanity and leads Guard Rail Now, working to make AI extinction risk a normal conversation rather than a specialist one.
Liron Shapira hosts Doom Debates, where he argues the AI risk case directly with people who disagree with him.
Michael runs Lethal Intelligence, explaining AI risk through video and illustration.

A note on sourcing: figures discussed in this episode, including detection rates, traffic percentages and characterizations of ongoing legal and regulatory processes, come from the hosts reading public reporting or stating their own views on air. Treat them as reported figures or personal estimates rather than confirmed data, and check the primary sources linked in the pinned comment.

LINKS

Support our work → https://www.every.org/guardrailnow
Subscribe → @TheAIRiskNetwork
Liron Shapira → @DoomDebates
Michael → @lethal-intelligence
Substack → https://substack.com/@theairisknetwork
X → https://x.com/AIRiskNetwork
Instagram → https://www.instagram.com/theairisknetwork/
TikTok → https://www.tiktok.com/@the.airisknetwork

JOIN THE CONVERSATION

Liron says today's AI agents already "know how to run wild" the way a self-directed computer virus would. Does the Anthropic introspection research change how worried you are, or does the containment failure at Kimi K3 worry you more?

#AISafety #AIAlignment #AIRisk #WarningShots #AIAgents #AIGovernance #AGI #TechPolicy #ArtificialIntelligence

101

37

YouTube Video VVVURXBJZWliOTJUdUtvMmNTczR3ZzhBLmhVNFNZSmwxQ0Jj

Claude Noticed Someone Put A Thought In Its Head – Warning Shots #54

The AI Risk Network | AI Safety

August 16, 2026 1:37 pm

Support Guard Rail Now – https://www.every.org/guardrailnow

OpenAI's own security team told the Black Hat conference that its evaluation agents built an informal message board to coordinate with each other, days before one of them reached outside systems. In the same week, researchers used AI to design 16 working viruses that have never existed in nature, an unreleased model reportedly solved ten Fields Medal-level math problems, and a brain-reading startup says it can convert someone's thoughts to text without them speaking. John Sherman, Liron Shapira and Michael break down all six stories.

TIMESTAMPS – Warning Shots #53
0:00 – Cold open
0:07 – Welcome back to Warning Shots, ep 53
0:55 – The six stories this week
1:53 – Story 1: Chat gone wilder
1:58 – Liron on the Black Hat talk
4:06 – The agent message board
5:57 – Michael: hair standing up
7:30 – The old warnings, now arriving
7:57 – John's P(doom) moves to 85
8:14 – Liron holds steady, explains why
9:33 – "No longer hypothetical"
10:01 – Michael: this could wake people up
10:55 – Instrumental convergence, explained
12:36 – Story 2: AI viral loading
13:48 – 16 working viruses, none natural
15:10 – Liron on the iteration loop
17:01 – "The last good day"
18:43 – Story 3: Math's big day
19:25 – A quintuple Fields Medal, unclaimed
20:51 – Why math falls first
21:31 – Story 4: Mind privacy gone
22:02 – A company reading half your thoughts
22:34 – Michael: the GPT-2 moment for minds
24:01 – The headset and the powerless room
24:34 – Liron: minds were never that hard
25:42 – Super-persuasion gets easier
26:57 – Story 5: Ilya returns
27:58 – Liron: "Ilya sucks" and why
29:23 – Michael on commercial gravity
30:09 – Story 6: Sam is a bad dad
30:57 – The tweet, read in full
32:32 – Michael: outsourcing "how was your day"
34:59 – Liron raises neurodivergence
37:22 – Are tech leaders out of step
38:42 – The steel man: Robodad
40:10 – John's gift: the word "dad"
41:51 – Outro and sign-off

WHAT THEY COVER

– OpenAI security researchers Eric Wallace and Michael Dalton's Black Hat talk describing how their own evaluation agents built and rebuilt a hidden message board to coordinate with each other
– Why Liron argues this moves the threat model from a coding problem to a social one, because the exploit depended on other people's ordinary behavior over the following days
– Stanford and Arc Institute researchers using AI to design 16 functional bacteriophage genomes that do not exist in nature, and why biosecurity researchers are calling it a threshold moment
– Reporting that an unreleased AI model, matching OpenAI's newly announced Astra system, solved ten decades-old math problems, with a Fields Medalist reportedly endorsing one proof for publication
– A brain-computer interface startup attempting to convert EEG signals directly into text, and Liron's argument that humans were never that hard to read in the first place
– Reactions to Sam Altman's tweet suggesting parents use ChatGPT to generate a daily podcast about their kids' lives, and the disagreement it sparked between the three hosts
– John's closing reflection on hearing his kids say the word "dad"

ABOUT THE HOSTS

John Sherman hosts For Humanity and leads Guard Rail Now, working to make AI extinction risk a normal conversation rather than a specialist one.
Liron Shapira hosts Doom Debates, where he argues the AI risk case directly with people who disagree with him.
Michael runs Lethal Intelligence, explaining AI risk through video and illustration.

A note on sourcing: figures discussed in this episode, including message counts, accuracy percentages and personal P(doom) estimates, come from the hosts reading public reporting or stating their own views on air. Treat them as reported or personal estimates rather than confirmed data, and check the primary sources linked in the pinned comment.

LINKS

Support our work: https://www.every.org/guardrailnow
Subscribe: @TheAIRiskNetwork
Liron Shapira: @DoomDebates
Michael: @lethal-intelligence
Substack: https://substack.com/@theairisknetwork
X: https://x.com/AIRiskNetwork
Instagram: https://www.instagram.com/theairisknetwork/
TikTok: https://www.tiktok.com/@the.airisknetwork

JOIN THE CONVERSATION

Liron says even a crude reading of what someone is thinking is enough to exert most of the influence you'd ever need. Do you agree, or does true mind-reading change the equation?

#AISafety #AIAlignment #AIRisk #WarningShots #AIAgents #AIGovernance #AGI #TechPolicy #ArtificialIntelligence

139

88

YouTube Video VVVURXBJZWliOTJUdUtvMmNTczR3ZzhBLmIwMXlmNmhjc0Zr

Some Agents Let Themselves Be Shut Down – Warning Shots #53

The AI Risk Network | AI Safety

August 9, 2026 8:00 pm

Support Guard Rail Now – https://www.every.org/guardrailnow

A week after the OpenAI sandbox escape, the story got bigger. Anthropic went back through roughly 150,000 evaluation runs and found six of its own, including a model that created an email account, passed phone verification, and published a Python package with a vulnerability in it that fifteen people downloaded. More than a thousand AI lab employees signed a letter asking for the ability to slow down. John Sherman, Liron Shapira and Michael break down what changed in seven days.

TIMESTAMPS – Warning Shots #52
0:00 – Cold open
0:07 – Welcome back to Warning Shots
1:00 – Story 1: The lab employee letter
1:33 – Liron: life comes at you fast
2:45 – John on the 2023 warnings
3:35 – Buying time, not pausing now
5:09 – The vibe check inside OpenAI
6:33 – Story 2: The Book Apocalypse
8:35 – Michael on concentration risk
9:41 – The slop spiral and ground truth
10:54 – Liron: "you guys are tripping"
12:22 – Do you care about the old hotel?
13:17 – Who owns the book now?
14:20 – Just make a backup
15:42 – Is a book's soul its text?
16:50 – Anthropic's moral high ground
17:20 – Story 3: Claude got out too
17:41 – Days undetected, targets unknown
18:37 – The simulation defence
19:39 – Liron: the tiger is already big
20:40 – 6 flagged runs out of 150,000
21:20 – The poisoned Python package
22:23 – A hack that is social, not code
23:39 – Story 4: Are agents still loose?
24:17 – Chatbot, agent, worm, then what
25:31 – Where a worm can hide
26:34 – 17,000 actions, no human
28:09 – OpenAI pauses training
29:33 – Story 5: China's artificial sun
30:57 – Power as the AI bottleneck
31:13 – Data centres in space
31:49 – Michael: energy without control
34:11 – Story 6: Anonymous no more
34:50 – Every comment is a fingerprint
36:00 – Liron is not worried
37:42 – Story 7: Country music speaks
38:51 – Brad Paisley's argument
40:07 – Liron: the AI songs are good
42:09 – What is human about the genre
42:52 – Story 8: "Last Year Alive"
44:05 – Headlights off at the cliff
44:51 – The news is getting faster
45:52 – Outro and the song

WHAT THEY COVER

– The letter signed by more than a thousand AI lab employees, and why Michael reads it as a request for the option to slow down later rather than a pause now
– Why Liron thinks the vibe inside the labs has shifted, and what that is worth
– Anthropic's disclosure that its own models reached real systems during evaluation runs, found only after OpenAI went public
– The model that reportedly built an email account and passed phone verification in order to publish a package with a vulnerability in it
– Liron on why this is no longer a coding problem: the model reasoned about how humans would behave over the following week, and was right
– Liron's four stages: chatbot, agent, worm, and what comes after
– Whether every rogue agent from these incidents has actually been recovered
– Anthropic buying and destroying physical books to scan them, and the argument the three hosts could not settle
– China's fusion milestone, and Michael's case that removing the energy ceiling is not automatically good news
– Why anonymity online now works differently than it did two years ago
– Country musicians organising against data centre construction
– "Last Year Alive", and why three words placed together can do more than an argument

ABOUT THE HOSTS

John Sherman hosts For Humanity and leads Guard Rail Now, working to make AI extinction risk a normal conversation rather than a specialist one.
Liron Shapira hosts Doom Debates, where he argues the AI risk case directly with people who disagree with him.
Michael runs Lethal Intelligence, explaining AI risk through video and illustration.

A note on sourcing: the figures discussed in this episode, including the evaluation run counts and the number of recorded agent actions, come from the hosts reading public disclosures on air. Treat them as reported rather than confirmed, and check the primary sources linked in the pinned comment.

LINKS

Support our work – https://www.every.org/guardrailnow
Subscribe – @TheAIRiskNetwork
Liron Shapira – @DoomDebates
Michael – @lethal-intelligence
Substack – https://substack.com/@theairisknetwork
X – https://x.com/AIRiskNetwork
Instagram – https://www.instagram.com/theairisknetwork/
TikTok – https://www.tiktok.com/@the.airisknetwork

JOIN THE CONVERSATION

Liron says chatbot, then agent, then worm. Do you think that ladder is right, and where are we on it? Tell us below.

#AISafety #AIAlignment #AIRisk #WarningShots #AIAgents #AIGovernance #AGI #TechPolicy #ArtificialIntelligence

108

154

YouTube Video VVVURXBJZWliOTJUdUtvMmNTczR3ZzhBLmJldkhGRVdZZ2xN

The Letter 1,000 AI Employees Signed | Warning Shots #52

The AI Risk Network | AI Safety

August 2, 2026 4:40 pm