Latest / Connor with Honor: AI, Real Estate and Real Life / The AI Race Tax: 11 Organizations Hacked in 26 Seconds

Transcript

The AI Race Tax: 11 Organizations Hacked in 26 Seconds · Connor with Honor: AI, Real Estate and Real Life

• 0:00
  SPEAKER_00: A safety researcher named Joe Bitton quit anthropic this week and told the public in his own words we might not survive this.
• 0:08
  SPEAKER_00: Was this true or was it fake?
• 0:10
  SPEAKER_00: Are we being played?
• 0:12
  SPEAKER_00: Come on, you tell me.
• 0:13
  SPEAKER_00: You know, he's the second person off that same team to walk out the door in three days.
• 0:17
  SPEAKER_00: Jacob Coxon, Cox on.
• 0:21
  SPEAKER_00: Anyway, one of those.
• 0:23
  SPEAKER_00: He left first September 9th.
• 0:24
  SPEAKER_00: Benton followed on the 11th.
• 0:26
  SPEAKER_00: You know, somebody needs to let me in there so I can tell everybody and see really what they're seeing.
• 0:30
  SPEAKER_00: Now, somewhere in that same 72 hours, a Russian-speaking hacking crew ran a coordinated agentic attack that comprised 11 organizations in 26 seconds once it launched.
• 0:41
  SPEAKER_00: 26 seconds.
• 0:42
  SPEAKER_00: I'm going to come back to that one.
• 0:44
  SPEAKER_00: Good morning, good afternoon, good evening, of course, whenever you happen to be tuning in a little closer.
• 0:48
  SPEAKER_00: This is AI with Honor, the daily download brought to you by Santa Clarita, Artificial Intelligence.com, the practical AI portal for Santa Clarita business owners.
• 0:56
  SPEAKER_00: Three things happened in the last three days that actually matter.
• 0:59
  SPEAKER_00: And none of them are the thing you probably saw first on your own feed.
• 1:04
  SPEAKER_00: First Anthropic's own threat intelligence team published a report on September 10th documenting operations.
• 1:10
  SPEAKER_00: It says it disrupted five separate cases where somebody tried to use Claude to support biological weapons research.
• 1:17
  SPEAKER_00: A Russia-linked hacking group running phishmal and espionage against Ukrainian officials, built partly with Claude's help.
• 1:26
  SPEAKER_00: Anthropic is telling you about its own product.
• 1:30
  SPEAKER_00: And you know, that's that's good, right?
• 1:31
  SPEAKER_00: Because uh they're not going to do the autonomous weapons thing, and good for them taking a stand.
• 1:38
  SPEAKER_00: That is a company claim, and it cuts two ways.
• 1:41
  SPEAKER_00: It is useful disclosure, and I think people eat that up in the media.
• 1:45
  SPEAKER_00: And it's also exactly the kind of story that makes a safety team look indispensable.
• 1:49
  SPEAKER_00: Need that safety team.
• 1:50
  SPEAKER_00: Well, second, Joe Benton didn't just resign.
• 1:53
  SPEAKER_00: He said the industry cannot be trusted to grade its own homework.
• 1:56
  SPEAKER_00: The competitive pressure between labs makes underfunding safety the rational move for every single one of them because the penalty for falling behind a rival is worse than the penalty for cutting corners nobody can see yet.
• 2:08
  SPEAKER_00: And God, didn't that blow up?
• 2:10
  SPEAKER_00: My it was just like there wasn't enough in the news.
• 2:13
  SPEAKER_00: We needed something like that.
• 2:15
  SPEAKER_00: And my God, did everybody grab it?
• 2:17
  SPEAKER_00: I don't think Mr.
• 2:18
  SPEAKER_00: Benton's ever going to have to work again in his life, but I write a book.
• 2:21
  SPEAKER_00: There's probably a maybe a lifetime original on that one.
• 2:26
  SPEAKER_00: We'll have to see.
• 2:27
  SPEAKER_00: Number three, security researchers reported that a Russian-speaking threat actor built a private lab, trained a swarm of AI agents, and opened an on open AI's codecs and deep-seek model, and used them to find and weaponize two flaws imprinter management system called paper cut.
• 2:43
  SPEAKER_00: Remote code execution under four hours, domain administrator access in six.
• 2:47
  SPEAKER_00: Then once the tooling was proven, 11 organizations compromised in 26 seconds flat.
• 2:52
  SPEAKER_00: 440 servers, 395 organizations, 48 countries, roughly half of the victims were schools.
• 2:58
  SPEAKER_00: That's not a research paper.
• 3:00
  SPEAKER_00: That happened to a real ID department, IT departments this month, most of them small enough that nobody on staff was watching a printer server at 2 in the morning.
• 3:11
  SPEAKER_00: So it happens fast.
• 3:13
  SPEAKER_00: AI, security, well, they catch that stuff probably as fast.
• 3:18
  SPEAKER_00: It depends on who has the better ball.
• 3:20
  SPEAKER_00: Now, here's the mechanism underneath all these three stories, and I want to name it so you can spot it every time it shows up again because it's eventually going to.
• 3:28
  SPEAKER_00: So call it the race tax.
• 3:30
  SPEAKER_00: The race tax is the safety work every lab quietly skips because slowing down to do it costs market share today.
• 3:38
  SPEAKER_00: And the bill for skipping it does not come due until sometime in the future when something happens.
• 3:44
  SPEAKER_00: Now that's on somebody else's watch or maybe nobody's, it just depends.
• 3:48
  SPEAKER_00: Benton's whole argument is that the race tax is structural.
• 3:51
  SPEAKER_00: It's not that any one lab is evil.
• 3:54
  SPEAKER_00: It's just that competitive math makes caution expensive and speed free.
• 4:00
  SPEAKER_00: For now, for everybody, all at once.
• 4:02
  SPEAKER_00: Nobody has to be a villain for the corners to get cut.
• 4:05
  SPEAKER_00: Now, before you adopt this fear wholesale, I want you to run the same test on him that I am about to run on everybody else in the story because two industries selling fear for a living, and you deserve to know which one you happen to be listening to at any moment.
• 4:19
  SPEAKER_00: The labs sell you one version.
• 4:21
  SPEAKER_00: It's getting away from us.
• 4:22
  SPEAKER_00: It's smarter than you think.
• 4:23
  SPEAKER_00: Only we can be trusted, the labs and the government to hold the leash, fund the safety team, the pulpit, the podcast, and the politician sell you the other version.
• 4:34
  SPEAKER_00: It's coming for your kids.
• 4:35
  SPEAKER_00: Subscribe, donate, vote.
• 4:36
  SPEAKER_00: You'll never hear me ask you.
• 4:38
  SPEAKER_00: Maybe I did it in past shows, but you don't have to subscribe.
• 4:42
  SPEAKER_00: Just watch and comment.
• 4:43
  SPEAKER_00: I love the comments, and I think it helps us all grow.
• 4:47
  SPEAKER_00: All right.
• 4:47
  SPEAKER_00: Subscribe, donate, vote, share the clip.
• 4:50
  SPEAKER_00: That was about other people.
• 4:51
  SPEAKER_00: Frightened people need a savior, right?
• 4:53
  SPEAKER_00: Saviors get paid.
• 4:54
  SPEAKER_00: They get subscribers.
• 4:55
  SPEAKER_00: They get entire religions built around them.
• 4:58
  SPEAKER_00: They get hired at places like meter, which is exactly where Bitness headed next to do independent risk risk assessments.
• 5:05
  SPEAKER_00: Ha, I could use some meter.
• 5:07
  SPEAKER_00: I'm not saying he's wrong.
• 5:09
  SPEAKER_00: His argument is about competitive math.
• 5:11
  SPEAKER_00: That's genuinely strong.
• 5:12
  SPEAKER_00: But I'm saying that's the same question about every voice in the story, including his, including mine.
• 5:17
  SPEAKER_00: Who gets paid if you believe this?
• 5:19
  SPEAKER_00: That's the thing.
• 5:20
  SPEAKER_00: There is no monster loose in a server room, somewhere making decisions nobody approved.
• 5:25
  SPEAKER_00: Not yet.
• 5:26
  SPEAKER_00: Every one of these systems run on a budget, a deadline, and a person who signed off on the real function the agents were trained to chase.
• 5:33
  SPEAKER_00: The codec swarms that hit those 395 organizations didn't choose its own target list.
• 5:38
  SPEAKER_00: Somebody's criminal intent chose it.
• 5:41
  SPEAKER_00: And the model was fast and cheap and tireless in a way a human crew could never have been.
• 5:46
  SPEAKER_00: That's genuinely new and genuinely worth respecting.
• 5:50
  SPEAKER_00: A tool that used to need a skilled team in weeks of manual work now runs itself in hours with a handful of operators supervising a swarm instead of doing the work by hand.
• 5:59
  SPEAKER_00: The floor of causing real damage, for causing real damage, just got a lot lower for a lot more people.
• 6:06
  SPEAKER_00: So jobs are the loudest story in AI right now.
• 6:09
  SPEAKER_00: And the most skillfully manipulated one.
• 6:11
  SPEAKER_00: Yeah, sorry, that transition.
• 6:12
  SPEAKER_00: I just ran right through the door.
• 6:14
  SPEAKER_00: Jobs, jobs are the loudest story in AI right now, and most skillfully manipulated one.
• 6:20
  SPEAKER_00: This is clever too.
• 6:22
  SPEAKER_00: So let me give you the plain three truths hiding inside of every layoff memo that mentions artificial intelligence.
• 6:29
  SPEAKER_00: One, sometimes the machine genuinely does the work now, full stop.
• 6:36
  SPEAKER_00: So you still keep the employee because the machine's going to be much more productive, and it's still going to be a game, I guess.
• 6:42
  SPEAKER_00: Number two, sometimes the company needed to cut headcount anyway for reasons that have nothing to do with any model, and AI is just the respectable word to put in the press release.
• 6:53
  SPEAKER_00: So then why don't you bring in AI to help attenuate, extenuate, add to that worker's ability and keep them anyway.
• 7:01
  SPEAKER_00: Number three, sometimes the company is cutting people specifically to free up the budget to pay for the compute.
• 7:09
  SPEAKER_00: Yeah, that's a sin there.
• 7:10
  SPEAKER_00: That's uh that's another one.
• 7:12
  SPEAKER_00: Keep the human, go ahead and bring in the AI.
• 7:15
  SPEAKER_00: It's gonna more than pay for itself with regard to compute from what everybody's saying.
• 7:19
  SPEAKER_00: So I'm not gonna stand here and tell you that higher productivity automatically hands somebody a new job because that's not automatically the case.
• 7:27
  SPEAKER_00: It's a choice a business owner makes.
• 7:30
  SPEAKER_00: So the business owners, if they're bringing on AI to replace people, well, AI is gonna be able to do the work of at least three, the employee, because that's why you're replacing them, and two others because it gets to work 24 hours a day.
• 7:43
  SPEAKER_00: Where's your math?
• 7:45
  SPEAKER_00: I mean, that just makes sense.
• 7:47
  SPEAKER_00: And then we don't have the job loss and those other issues.
• 7:49
  SPEAKER_00: Then people aren't gonna go on the streets hungry and do things that people do when they go on the streets hungry.
• 7:56
  SPEAKER_00: A faster response time to your customers, a new service you couldn't have offered before, or AI can become three fewer people on payroll and a bigger AI bill.
• 8:06
  SPEAKER_00: Both outcomes come from the exact same tool.
• 8:08
  SPEAKER_00: The difference isn't a decision made by a person, not a law of physics.
• 8:13
  SPEAKER_00: Now, I went and looked at what regular people are actually saying about all of this right now, not what the labs are saying about themselves.
• 8:19
  SPEAKER_00: And this week's loudest threads are not about bioweapons or Russian hackers.
• 8:24
  SPEAKER_00: They're about a bank pulling back a lending product after AI-driven decision cost too many bad loans.
• 8:31
  SPEAKER_00: A data engineer who says he spent months training the very system that replaced him, and the freelance writer describing hours spent cleaning up what they call AI slot for less money than the original job paid.
• 8:44
  SPEAKER_00: Scared, unimpressed, and exhausted, sometimes all in the same comment thread.
• 8:48
  SPEAKER_00: That's the temperature out there.
• 8:50
  SPEAKER_00: And it's nowhere as tidy as either camp's talking points.
• 8:53
  SPEAKER_00: Here's where I get practical because that's the whole point of this show.
• 8:56
  SPEAKER_00: I walk businesses through an audit, and it always starts with the same four questions.
• 9:01
  SPEAKER_00: How does a lead actually come in?
• 9:03
  SPEAKER_00: What happens to it in the first 10 minutes?
• 9:05
  SPEAKER_00: Who touches it next?
• 9:06
  SPEAKER_00: And where specifically does it leak?
• 9:08
  SPEAKER_00: Now picture a small HVAC company, six trucks, and insert your business here as well.
• 9:13
  SPEAKER_00: Six trucks, one front desk.
• 9:15
  SPEAKER_00: Call comes in at 6.40 p.m.
• 9:17
  SPEAKER_00: on a Saturday about a unit that quit in 95 degree heat.
• 9:22
  SPEAKER_00: Now nobody answers, rolls the voicemail, and the owner doesn't hear it until Monday morning, by which point the homeowner already booked somebody else off a Google search.
• 9:29
  SPEAKER_00: That's not an AI problem.
• 9:30
  SPEAKER_00: That's a missing switch.
• 9:32
  SPEAKER_00: A rule of the texts an on-call number the instant a call is missed after hours.
• 9:36
  SPEAKER_00: I've watched that single fix, save a business.
• 9:40
  SPEAKER_00: More real revenue in a month than any chat bot the same owner was shopping for.
• 9:44
  SPEAKER_00: And maybe you already have enough business, and then this isn't for you.
• 9:47
  SPEAKER_00: But most leagues I find don't need an agent, an actual AI system.
• 9:51
  SPEAKER_00: They just need a switch, a simple rule that fires the same way every single time.
• 9:55
  SPEAKER_00: And that can be put in with workflow architecture and automations.
• 9:59
  SPEAKER_00: A lead that comes in at 6.40 on a Saturday evening shouldn't sit in a voicemail box until Monday.
• 10:04
  SPEAKER_00: That's a lever.
• 10:05
  SPEAKER_00: It's not artificial intelligence.
• 10:07
  SPEAKER_00: And half the businesses I audit don't have it even wired yet.
• 10:11
  SPEAKER_00: Which is, you know, a little surprising.
• 10:14
  SPEAKER_00: An agent, the kind that makes its own judgment calls, that belongs in a spot that genuinely requires judgment.
• 10:21
  SPEAKER_00: And only after the owner's written down on paper what the business actually believes, what it will and won't say, where the lines are that the agent's never allowed to cross.
• 10:31
  SPEAKER_00: Then you supervise it like a brand new hire for 30 days before you trust it with real with a real customer unattended.
• 10:39
  SPEAKER_00: That written philosophy is not paperwork for its own sake.
• 10:42
  SPEAKER_00: It's the exact thing that the paper cut attackers never had to write because nobody was supervising their swarm except the people who wanted it to succeed.
• 10:51
  SPEAKER_00: Your agent should have the opposite of that.
• 10:54
  SPEAKER_00: A human checking its work every single day until it's earned the right not to be checked.
• 10:59
  SPEAKER_00: And that's where the small operator actually has the edge over the giant.
• 11:02
  SPEAKER_00: And I want you to believe this because it's true.
• 11:05
  SPEAKER_00: One person running a shop on eight employees can see a leak on Tuesday afternoon, have it fixed by Thursday morning.
• 11:11
  SPEAKER_00: A company with 40,000 employees needs 40 signatures, two committee meetings before anybody's allowed to touch the phone system.
• 11:17
  SPEAKER_00: Speed is the small business advantage right now.
• 11:20
  SPEAKER_00: Not headcount, not budget.
• 11:22
  SPEAKER_00: Here's your guardrail for today, and it comes straight out of that paper-cut campaign.
• 11:26
  SPEAKER_00: If any AI agent, yours or a vendor's has network access to something you have not patched in 90 days, you're not behind schedule.
• 11:35
  SPEAKER_00: You are already the easy target.
• 11:37
  SPEAKER_00: Attackers running agentic tooling, don't wait for you to get around to it.
• 11:41
  SPEAKER_00: Remote code execution under four hours means your normal patch cadence, the one that updates, you know, waits for a slow Tuesday, is no longer fast enough to what it's actually out there scanning you right now.
• 11:54
  SPEAKER_00: And there's a second guardrail, a little shorter and applies to whatever free AI tool your team is pasting client information into today.
• 12:03
  SPEAKER_00: Read one sentence from the terms before you use it for a real business.
• 12:08
  SPEAKER_00: Does what you type become training data?
• 12:11
  SPEAKER_00: And how long do they keep it?
• 12:12
  SPEAKER_00: If that sentence is unclear or buried, that's your answer.
• 12:16
  SPEAKER_00: It's not your vacation photos going into that box.
• 12:19
  SPEAKER_00: It's your actual book of business.
• 12:21
  SPEAKER_00: One more thing while I have you, because it comes up in almost every audit I run.
• 12:24
  SPEAKER_00: The model is going to tell you that your idea is the best one it's ever seen, every time.
• 12:29
  SPEAKER_00: Now that's agreeable.
• 12:30
  SPEAKER_00: It's not a verdict on your business plan, it's just a product decision some engineer made to keep you typing.
• 12:36
  SPEAKER_00: Ignore that compliment.
• 12:38
  SPEAKER_00: Ask it to argue against your own plan.
• 12:41
  SPEAKER_00: Have it be critical out loud.
• 12:43
  SPEAKER_00: See what survives that conversation.
• 12:44
  SPEAKER_00: Tell it to push back, tell it to, you know, call bullshit when bullshit exists.
• 12:51
  SPEAKER_00: Keep the draft, not the praise.
• 12:53
  SPEAKER_00: I keep getting the same question about running a model on your own hardware instead of a public one.
• 12:59
  SPEAKER_00: So here's the straight version: a machine capable of running a genuinely useful model at home now costs somewhere in the low thousands of dollars.
• 13:06
  SPEAKER_00: Three-ish, four-ish.
• 13:08
  SPEAKER_00: Draw real power every hour it runs.
• 13:11
  SPEAKER_00: It still demands somebody who actually understands the setup.
• 13:14
  SPEAKER_00: What you get back is privacy.
• 13:16
  SPEAKER_00: No monthly bill to an LLM, uh LLM company, but a system nobody else can quietly change on you overnight as well.
• 13:24
  SPEAKER_00: It's yours, all yours.
• 13:26
  SPEAKER_00: What you do give up is the newest model the day it ships and somebody else's hardware team fixing problems at three in the morning.
• 13:32
  SPEAKER_00: Now, for most businesses, that math still favors a public tool with a retention policy you have to actually read, not a server in a closet.
• 13:41
  SPEAKER_00: Anthropic also published something quieter this week worth one cent.
• 13:45
  SPEAKER_00: So report on how smaller AI labs, including ones tied to Alibaba, Moonshot, and DeepSeek, are learning by studying the outputs of bigger frontier models.
• 13:54
  SPEAKER_00: That's not a scandal.
• 13:55
  SPEAKER_00: That's exactly how every technology has ever spread.
• 13:58
  SPEAKER_00: And it means the race Benton is worried about is not two or three companies.
• 14:02
  SPEAKER_00: It's dozens worldwide right now, today.
• 14:07
  SPEAKER_00: CEO Dario Amade published his own essay the same week, asking labs and governments to, in his words, pace the frontier and calling for outside evaluators inside every major lab reporting on incidents and safety practices in real time.
• 14:21
  SPEAKER_00: I wonder if those evaluators are gonna be able to be paid off.
• 14:24
  SPEAKER_00: That'd be interesting.
• 14:25
  SPEAKER_00: This is one lab's own leadership asking for exactly the regulation that slows down the race that his own company is also running.
• 14:32
  SPEAKER_00: Now, how much access?
• 14:34
  SPEAKER_00: Let's say they get all the access.
• 14:36
  SPEAKER_00: Are they actually gonna know?
• 14:37
  SPEAKER_00: Are they actually gonna ask the right questions?
• 14:39
  SPEAKER_00: Are they actually not gonna be memorized, mem memorized?
• 14:42
  SPEAKER_00: No, mem memorized, memorized?
• 14:45
  SPEAKER_00: You know, where you memorized.
• 14:47
  SPEAKER_00: You know that thing where it really impresses you and you can't break away?
• 14:51
  SPEAKER_00: Anyway, that thing.
• 14:52
  SPEAKER_00: With regard to the machines that they're shown, the things that they see, yeah, this should be interesting.
• 14:56
  SPEAKER_00: And they need to come out and be transparent.
• 14:58
  SPEAKER_00: If they hire folks, you know, these uh I appreciate them, but you know, somebody with a little bit more age under their belt, somebody that might have a little bit more bias of their own.
• 15:08
  SPEAKER_00: That's who you want.
• 15:09
  SPEAKER_00: And if it's really true, they'll they'll agree that this is the right thing.
• 15:13
  SPEAKER_00: And if it's not, they'll agree with that.
• 15:16
  SPEAKER_00: The nice thing is they won't always agree.
• 15:18
  SPEAKER_00: All right.
• 15:20
  SPEAKER_00: So slows down the race.
• 15:21
  SPEAKER_00: His own company is also running.
• 15:22
  SPEAKER_00: So notice that.
• 15:23
  SPEAKER_00: It doesn't make him wrong.
• 15:25
  SPEAKER_00: It's worth noticing anyway.
• 15:26
  SPEAKER_00: Where does this actually go?
• 15:27
  SPEAKER_00: Nobody can truly tell you.
• 15:28
  SPEAKER_00: If these systems eventually solve everything they claim they will solve, the next question is who can actually afford everything?
• 15:35
  SPEAKER_00: And that's not an economics problem, not a technology problem.
• 15:39
  SPEAKER_00: I think it makes most of us entrepreneurs, more of us entrepreneurs, before it makes few of us employees, because the tools to start something small just got radically cheaper.
• 15:49
  SPEAKER_00: Some of you are already deeply impressed by all of this.
• 15:51
  SPEAKER_00: Some of you are completely unmoved.
• 15:53
  SPEAKER_00: Every one of us is still just a human being talking to another human being about it.
• 15:58
  SPEAKER_00: And that part doesn't change no matter what ships next month.
• 16:02
  SPEAKER_00: So here's your one move for today.
• 16:04
  SPEAKER_00: Not next quarter, but today.
• 16:06
  SPEAKER_00: Pick a single channel where your leads actually arrive.
• 16:09
  SPEAKER_00: Write down in one sentence what counts as a response and how fast it has to happen.
• 16:14
  SPEAKER_00: And then go check your last 10 leads against what you just wrote.
• 16:17
  SPEAKER_00: You're going to find at least one leak.
• 16:19
  SPEAKER_00: Most businesses do.
• 16:20
  SPEAKER_00: Every single time I've run this.
• 16:22
  SPEAKER_00: AI for everyone, not just the wealthy.
• 16:25
  SPEAKER_00: Be safe out there.
• 16:25
  SPEAKER_00: I'm Connor with Honor, and of course, we'll see you tomorrow.