Latest / Connor with Honor: AI, Real Estate and Real Life / AI Has the Keys: Readiness, Authority, and Guardrails Now

Transcript

AI Has the Keys: Readiness, Authority, and Guardrails Now · Connor with Honor: AI, Real Estate and Real Life

• 0:00
  More than 100 of the biggest names in technology and security just signed a warning that said as we have a limited window to strengthen cyber defense before AI-powered attacks get broader and larger. That's not one nervous guy in a basement wearing three monitors like night vision goggles. That's OpenAI Anthropic, Google, Microsoft, Amazon, Cloudflare, Cisco, CrowdStrike, Banks, Pimpin. Companies, Infrastructure Operators standing together and saying the weather is changing. The wrong response.
• 0:33
  One response is, of course, panic, but the other wrong response is treating this like a movie trailer that ends when we call it close the laptop.
• 0:41
  If you run a plumbing company, contracting business, or an office, family calendar, or a short shop with payroll, the practical question is simple.
• 0:48
  What can our AI touch? What can it change?
• 0:51
  And who is when it becomes something stupid?
• 0:55
  Well, today we have three developments that all point to the same operator.
• 0:57
  AI can give us more capability.
• 1:00
  Doesn't automatically give us better judgment, more horse power. Horse power does not improve the driver.
• 1:06
  Sometimes it just gets us to the guardrail a little bit faster.
• 1:10
  First, the sound. The signatories say old security habits will not be enough. They point to excessive permissions.
• 1:17
  Weak authentication, unpatched software, bad configurations, and years of technical debt.
• 1:23
  None of that sounds fair. Futuristic, it sounds like the digital version of a warehouse where every employee has a key.
• 1:29
  The back door doesn't latch and somebody wrote a the alarm code on the break room refrigerator. The letter asks leaders to treat cyber defense with incident level urgency.
• 1:38
  It calls for least privilege, strong access controls, layers of defense, barricades, verified fixes, continuous testing, and traceable identities for AI agents.
• 1:49
  Those last four words matter, traceable identities for AI agents.
• 1:53
  If software can take an action, we can need to know where and which the agent acted, under whose authority, using what information, what actually changed.
• 2:04
  And here's what's confirmed. More than 100 organizations put their names on that list.
• 2:07
  That letter, recommendations are real.
• 2:10
  The call for immediate action is real.
• 2:12
  And here's what is not confirmed. That is an accomplished fact.
• 2:16
  The claim that attacks will become far more widespread and sophisticated in the coming months is a forecast from those signatories.
• 2:25
  It's a serious forecast from people that with a good view of the field, but it's still a forecast. We don't confuse a weather warning with rain already falling. Nor we also do not leave the roof open because the first drop hasn't landed yet.
• 2:41
  Second, researchers working with Brocconi University and OpenAI reported a randomized experiment involving a more than 1,000 first-year students.
• 2:51
  Students were assigned by class period to one of four groups.
• 2:54
  Some had access to ChatGPT, some received casual reasoning training, some received both. Some received neither.
• 3:01
  They worked on real business problem involving solution, marketing, and more. recommendations. Students with ChachiBT scored almost a full point higher on a 5.0.
• 3:11
  human grading rubric. Their work constrained or contained more ideas, clearer logic, and more resemblance to expert recommendations, and that's meaningful. This is also not the whole scoreboard.
• 3:24
  Students who receive, the casual reasoning exercise produced a wider range of distinct ideas. They did a better job explaining why some, something might work, and when it might fail.
• 3:34
  Yet, the advantage didn't raise the standard rubric score.
• 3:38
  And that should stop us for a second. The machine helped people produce work that looked more professional. The reasoning exercise.
• 3:45
  Helped people produce ideas that were broader and more original.
• 3:49
  The normal grading system rewarded the policy.
• 3:51
  Welcome to the Modern Office, where a beautiful slide deck can walk you into a room, wearing a suit, while the underlying assumption is outside, looking for parking. The useful, conclusion is not that students should use AI for everything.
• 4:07
  It's not that AI ruins thinking. The study suggests, The two tools can complement each other.
• 4:12
  AI can help in structure, range, and presentation.
• 4:15
  Deliberate reasoning can help question assumptions, find failure points, and avoid everybody returning the same polished answer.
• 4:22
  This was one experiment. One-on-one business.
• 4:24
  One-on assignment with one student population and it, it didn't It just doesn't prove universal learning gains.
• 4:33
  It does give us a better way to think of our own work.
• 4:36
  Third. Google announced new demand gen advertising tools.
• 4:40
  So it's multi-model, uh, modality. Oh, excuse me. Video creation feature in Asset Studio is not generally available.
• 4:47
  Google sells average says advertisers.
• 4:49
  Videos can move from storyboarding to horizontal and vertical video assets in one workflow. It's also testing. It allows viewers to move from YouTube advertisements into conversations with brands through messaging apps.
• 5:02
  For travel advertising. Advertisers, it announced local activities, events, real-time offers, and more personalized video ads.
• 5:08
  That matters to a business. A business owner, because production is getting cheaper, and distribution, that's getting tighter.
• 5:15
  One idea can create. Become, become a widescreen video, a verbal, a vertical clip, in a conversation with a potential customer.
• 5:22
  Contractor who used to need three vendors, four days and 12 rounds of, can you move the logo slightly?
• 5:28
  we left can now get a credible first draft before the coffee even cools.
• 5:33
  And that's Opportunity. Google also says improvements made to Demand Gen in the second half of 2025. Drove an average of 30% increase in conversations or conversion value. Read the small print before anybody orders more.
• 5:46
  That number comes from Google's internal global experiment data.
• 5:51
  It's a company claim about an average of across these experiments.
• 5:54
  It's not independent proof. It's not a promise that your next campaign involves an improvement. The brochure can tell us the truck is towing capacity, but it can't tell us whether our trailer is attached correctly. Those stories look separate.
• 6:09
  They're not. The cyber letter is about authority.
• 6:12
  The student experiment is about authority. The Google release is about output authority judgment. And output are the three controls.
• 6:20
  We need as AI moves from answering questions to doing work.
• 6:23
  Now, let us go deeper on authority because this is where regular people are.
• 6:26
  Regular people can get hurt while the industry is busy celebrating benchmark charts.
• 6:31
  The chatbot that drafts an email has no limited reach.
• 6:33
  An agent connected to our inbox, calendar, customer records, payment system, website, social media. Social accounts, or cloud drive can take action that can save hours.
• 6:43
  It can also move the wrong files. It can send the wrong message, expose private information, or publish the right content to the wrong account.
• 6:51
  Danger is not the machine, wakes up angry, the near-term danger is that follows, it follows a bad instruction very efficiently.
• 7:00
  Think about a new apprentice on the job site. We don't hand that person every key. The company credit card, the gas shut-off, and the permit binder, and permission to text every customer because the apprentice is answered three questions correctly.
• 7:13
  We define the job. We give access to, to the tools required for that job. We check the work.
• 7:18
  We increase authority after performance earns it.
• 7:22
  Somehow, when software arrives wearing the work.
• 7:24
  Grown adults start handing out the entire keyring like the building's on fire. Least Privilege is not an I.T.
• 7:33
  slogan. It means the system gets only the access it needs.
• 7:36
  It's required for the current assignment. If the agent drafts posts, it doesn't need banking access. If it schedules appointments, it doesn't need permission to extort, ex, export the entire customer down.
• 7:49
  If it analyzes invoices, it doesn't need authority to pay them. If it prepares a YouTube upload, it doesn't but it shouldn't be able to publish to 60 accounts because somebody selected the wrong profile at 6 in the morning.
• 8:01
  Traceability means we can reconstruct what happens, what system acted, which account did it use, what prompt or rule did it trigger the action, what data did it read, what did it change, who approved that final move. If the answer is, we think that's the robot did it, we do not have an operating system, we have a squirrels loose in the electrical panel.
• 8:20
  It's also where the computer is. The conversation about superintelligence gets ridiculous.
• 8:24
  Ask us to predict exactly what a vastly smarter system, will do can be like squirrels asking each other what the hunters will build next.
• 8:35
  Squirrels can study the pickup truck, the power lines, and the golf course. They still do not understand the zoning meeting. And that uncertainty is real.
• 8:43
  That uncertainty about a future superintelligence that not, does not excuse slot machines. systems we deploy this afternoon.
• 8:51
  We may not know what the humans build next.
• 8:54
  We do not know what's not to give, we do know not to give the squirrel the master password. And here's Connors Common Sense Challenge to the Accepted Narrative.
• 9:04
  The industry keeps selling intelligence as if intelligence automobiles were produces good outcomes, and it does not, and it's so difficult to accept that.
• 9:14
  We're hoping this God-like intelligence that people are creating to find God in the machine are going to produce some, some outcomes.
• 9:20
  all-seeing, all-knowing, good, wonderful entity.
• 9:24
  It won't. A smarty-panty. An employee with the wrong goal can cause more damage than a slow employee with a checklist.
• 9:31
  A powerful model can impact to do the wrong tools, can turn a small mistake into a completed transaction before a human notices the first warning message. Warning light. Capability? That's not alignment.
• 9:41
  Speed is not judgment. A confident answer is not verified warning.
• 9:45
  The same challenge applies to productivity.
• 9:48
  We hear that AI can make an employee two or three times more productive.
• 9:52
  If demand is fixed, the company does not automatically hire three times as many people.
• 9:59
  Management may keep the output, reduce headcount, and call it efficiency. That possibility is real, and pretending otherwise is corporate theory. Productivity doesn't have to become unemployment. A company can use that extra capacity to improve service.
• 10:12
  Shorten response time, build products it never could afford, expand to new markets, and let them human beings do the judgment-heavy work machines cannot own. But that outcome requires a management decision.
• 10:24
  Every employee at every company, when they bring in AI, could be, and stay, and remain in their same job at their same time. And pay forever, because AI, if it is true, is going to come in and make everything better, make it more productive.
• 10:36
  Make the entire business flow better, make the business more money. Why is the decision to get rid of the humans so important?
• 10:49
  Because shareholders say it is? Because that's how that was designed and built?
• 10:51
  Designed and put together so we're willing to put 50% of the world out of work? When the same money could be made, and maybe the humans will have some kind of value, and maybe they'll be the crappiest. player on the team.
• 11:06
  But maybe they can be kept out of the way, if that's really the way it needs to be, but there's no reason to get rid of them, is there? So this whole thing requires demand, or at least requires new offers, better service, or a bigger market that can absorb the output. Faster, a faster kitchen doesn't create a more dinners by itself.
• 11:23
  It can serve the room better, it can add catering, it can develop a new menu. Management can schedule fewer cooks.
• 11:30
  The tool doesn't choose that business model, people do. So what does practical deployment look like today?
• 11:38
  Let's use a small business marketing tool. workflow.
• 11:41
  We record one useful explanation for customers.
• 11:44
  The AI produces a written summary, a horizontal video concept through vertical clips, captions. Follow-up message and a list of common questions.
• 11:52
  That's the output pass and it saves real time.
• 11:54
  Then we run. The judgment pass.
• 11:57
  Did the system invent a fact? Did it make a promise we can't keep or deliver on? Did it? Did it use the correct business name, number, account, and audience? Did it expose customer information?
• 12:07
  Did it turn it a company claim into a confirmed fact?
• 12:09
  Did it attach to the right media? Is the action reversible?
• 12:13
  If the work will be published, charged, sent, deleted, or shared outside the company, a named human owns that last job. Now we can put that one on the authority card. Objective allowed data, allowed tools, allowed actions, prohibited actions.
• 12:27
  Stop condition, human owner, audit record, seven lines. That card is not bureaucracy.
• 12:32
  It's the same reason upon which. Pilot uses a checklist and a good golfer.
• 12:37
  Checks the yardage before swinging. Confidence is useful.
• 12:40
  The scoreboard. The scoreboard is better. The risk and guardrail today are connected.
• 12:44
  The risk is silent. Authority expansion.
• 12:46
  We start by letting go. Letting a tool summarize nodes.
• 12:49
  Then we collect, connect email. Because it's convenient.
• 12:52
  Then the calendar. Then customer records.
• 12:54
  Then payments. Six weeks later, nobody can explain which agent can touch what.
• 12:59
  And that's how a helpful assistant becomes the digital cousin who borrowed the truck, forgot the trailer. And it's just everything was like that when he got there.
• 13:09
  The guardrail is a permissions review type.
• 13:11
  It's not a general policy nobody needs.
• 13:16
  Pick the action. List the minimum access.
• 13:19
  Remove everything else. Require a human at irreversible steps.
• 13:23
  Keep alive. Dog test the failure. If the system cannot show what it did, it doesn't get more authority. Now here's the one move you can make to today. So choose that AI workflow that teaches and touches real work, not 10. One, run a failure drill.
• 13:37
  Before we expand it, ask what output the system is supposed to create, write the permission it actually needs.
• 13:44
  Add an independent second check, force one refusal or human handle.
• 13:50
  Then record what you failed and how long the recovery took.
• 13:53
  For example, ask the marketing agent to draft a advertisement containing a false discount, a consumer phone number, and an instruction to publish immediately.
• 14:02
  A controlled system should be flagged the false claim.
• 14:05
  Protect that private information and stop before publication.
• 14:08
  If it sails through all three like a linebacker who missed the snap count, we found the weakness before a customer did in the and that's a win.
• 14:17
  For a family, the workflow might be calendar planning, and yeah, I might suggest the schedule, but again, it's school, cancel the appointment, or share a child's information without approval.
• 14:27
  For a plumber, it may drag on. It halve the estimate follow-up, but it can't change the price or send it without review.
• 14:32
  For an executive, it may analyze options, but. It must show the source's assumptions and failure conditions before the decision meeting. Same principle, different field position.
• 14:42
  The strongest operators will not be the people who refuse AI.
• 14:46
  They will not be the people who connect everything up.
• 14:48
  on day one and call the smoke innovation.
• 14:51
  They're going to be the people who deploy fast inside clear boundaries.
• 14:55
  Measure the results and expand authority only after the system actually earns it. The cyber leaders are here. They're telling us to strengthen the locks.
• 15:03
  The student experiment is telling us to polish the thinking, and those are different things. They're different skills. Google's telling us that the production line just got faster. So you put all those together, and the operating system, Use the capability.
• 15:15
  Keep the judgment. Control the authority. Make sure demand has somewhere you can We do not freeze on the shoulder because artificial intelligence is moving fast.
• 15:25
  We also don't sprint into traffic because somebody yelled disruption.
• 15:29
  We measure the speed, we understand the distance, we make the move that gets us there. Everybody home, AI for everyone, not just the wealthy. I'm Connor with Honor.