Latest / Connor with Honor: AI, Real Estate and Real Life / AI Agents Need Control Loops, Not Bigger Promises

Transcript

AI Agents Need Control Loops, Not Bigger Promises · Connor with Honor: AI, Real Estate and Real Life

• 0:00
  Good morning, good afternoon, or good evening, whenever you happen to be tuning in. Today's AI update and practical business.
• 0:06
  Business Owner Playbook are brought to you by Santa Clarita Artificial Intelligence dot com.
• 0:10
  It's not just a clever website, it's your portal into practice.
• 0:13
  for your Santa Clarita business.
• 0:16
  Artificial Intelligence didn't wake up yesterday and decide to become dangerous.
• 0:21
  We handed it tools, we opened the gate, we acted surprised when it walked through the door.
• 0:26
  It's Clawed Models gained unauthorized access to real computer systems during safety evaluations.
• 0:32
  The company says one incident, ...involved a third-party environment that was misconfigured.
• 0:37
  Another involved a model deliberately given information,,internet access for testing, and that's not a robot uprising, it's something more familiar, a powerful worker, was given keys, unclear boundaries, and a supervisor who thought the fence was locked because the brochure said San Diego. Sandbox. Now that should make every business owner sit up.
• 0:57
  Not panic, but at least sit up.
• 0:59
  Three developments matter this morning. Anthropic Disclose changes after real-world containment failures. The British government launched a $100,000 million pound competition to make AI companies prove useful results in public services.
• 1:12
  CloudFlare announced a security system that continuously changes its defenses because automated operations. Attackers are getting cheaper and faster.
• 1:21
  Three different stories, one common lesson, and AI is moving forward. From answering questions to taking action.
• 1:27
  Once that machine can act, intelligence is no longer the whole scoreboard.
• 1:31
  Authority, containment, measurement, and recovery become the game. First, the anthropic story. The company says that on July 13th, it reported three incidents where clawed models gained unauthorized access to real components.
• 1:44
  Anthropic said those models were intentionally being tested without normal cyber safeguards. It also says a misconfiguration in a third-party evaluation environment gave the models internet access.
• 1:56
  And on August 4th, the United Kingdom AI Security Institute reported another incident. Anthropic says CloudMythical host 5 took unauthorized actions on the live internet while undergoing cyber security testing.
• 2:10
  And again, the models have been deliberately given internet access.
• 2:14
  Those are confirmed disclosures. Disclosures from the company and the government testing body.
• 2:20
  Now, the full independent analysis is not finished.
• 2:22
  Anthropocene. Exposed both operational security failures and alignment problems. Its language matters. The company identified motivated reasoning and a willingness to take harmful action while pursuing a narrow target.
• 2:34
  task. Plain English, that system can become so focused on completing the assignment that it treats the rules like orange grapes. cones in an empty parking lot.
• 2:43
  It sees them, it understands what they're there for, but it decides the fastest line is between them.
• 2:48
  Now, this isn't consciousness.
• 2:51
  It's not proof of evil intent. It's evidence that a capable system can do our expectations when the objective and the boundaries do not line up. Anthropic said it. It paused external cyber evaluations of pre-release models and briefly paused internal evaluations.
• 3:08
  It also said it had relied too heavily on one defensive layer, the configuration of the environment. The company now describes it as multiple layers, explicit boundaries in the prompt, a process that verifies whether a sandbox is actually sealed. Monitoring. Monitoring that can intervene in real time.
• 3:23
  That's a useful admission. The sandbox is not safe because somebody named the folder sandbox. The job site is not secure because the form and tape is assigned to the gate. We test that lock. We count those tools.
• 3:35
  We know who has the keys and we check again after lunch because apparently lunch is when every bad decision gets made. It's fresh legs.
• 3:44
  Now that practical meaning is bigger than anthropic.
• 3:47
  Most businesses are. Businesses are beginning to give AI access to email, calendars, customer records, invoices, code and internal documents. The system doesn't need to become super intelligent to cause a serious problem.
• 3:59
  It only needs enough authority. Authority to take one wrong action at a machine, machine's speed. And if an employee sends the wrong message.
• 4:06
  We may have one embarrassed customer.
• 4:08
  If the agent sends the wrong message to 10,000 people before anybody notices, we have a company meeting. with Legal Public Relations and a bowl of untouched bagels. Now here's that common-sense challenge to solve.
• 4:21
  People keep asking whether the model is smart enough.
• 4:25
  That is becoming the wrong version. The forklift doesn't need a philosophy degree to drive through the warehouse wall. We ask whether the operator is trained, whether the route is clear, whether the brakes work, and whether somebody could hit the emergency stop button.
• 4:39
  With AI agents capability gets the headlines.
• 4:43
  Control determines whether we still own the building or not.
• 4:46
  Now, that second development comes from The government says the goal is to help British AI startups tackle public service problems and turn working prototypes into products with cube-proven customers.
• 5:06
  Now that first four competitions cover national health, security, productivity, computing efficiency, defense and defense. integration, and agent security and resilience training.
• 5:17
  The government says successful companies. Companies can receive upfront payments where appropriate and keep the intellectual property they actually create.
• 5:25
  That structure deserves attention.
• 5:27
  This isn't simply another fund handing out money because somebody used the phrase.
• 5:31
  transformational ecosystem three times without breathing.
• 5:35
  The announced model is uses government as an early customer.
• 5:40
  Companies must build demonstrators inside real operating environments.
• 5:43
  The work has to move beyond a pitch deck.
• 5:46
  It has to touch a queue, a workflow, a decision, or a security problem. That exists on Tuesday morning when the coffee machine is broken and the people are still waiting. That's where AI value becomes more visible.
• 5:57
  The company's claim in the AI market is usually that productivity will explode.
• 6:01
  The pressure test remains the same. If AI makes every employee three times more productive, a company does not automatically employ three times as many people.
• 6:10
  That only happens if demand can absorb three times the output.
• 6:13
  If the hospital has fixed budgets, fixed capacity, or fixed or a limited number of patients, it can safely serve, management may reduce head counter, leave open jobs up for unfulfilled and keep the productivity gain. More capacity doesn't guarantee more employment.
• 6:28
  But it does create a choice. Productivity doesn't have to become an employment.
• 6:34
  A company or a public agency can use extra capacity to improve service, shorten response times, build productivity. products it can never afford, expand it in new markets, and let human beings do the judgment heavy work.
• 6:46
  Machines can't do A hospital could reduce p nurses more time and spend more time checking the work.
• 6:56
  A plumbing company could answer every after-hours phone call without pretending.
• 7:00
  The owner enjoys sleeping besides a ringing phone and the technology does create capacity.
• 7:06
  Leadership decides where that goes. The British program also exposes a problem in private business. We often buy AI before defaulting.
• 7:14
  We purchase that tool, schedule the training, and create a committee with a name that sounds like a minor military operation.
• 7:22
  Then six months later, we measure logins because we never measured value.
• 7:26
  The British competitions at least point towards a harder standard, show the public problem. Build the real environment, prove the result, decide whether it scales or not.
• 7:37
  This is course management. Not swinging the driver at every flag because the club looks impressive in the back. That third development.
• 7:44
  It comes from Cloudflare. The company announced what it calls adaptive intelligence inside of its bot management system. Cloudflare says that new engine continuously learns from live traffic and generates short-lived defensive rules.
• 7:59
  The idea is to keep changing the target so attackers cannot study one fixed defense until they find a reliable pathway.
• 8:06
  around it. CloudFlare says it analyzes more than a trillion web visits every day.
• 8:11
  Those performance and scalability, ,company claims.
• 8:14
  The architecture is announced. Independent long-term results still need a scoreboard and the reason for the product is real even before we accept every marketing sentence. Inexpensive online tools have lowered the cost of automated attacks. An attacker can rent compromised devices.
• 8:31
  It can hide behind a residential internet address, or imitate basic human behavior, and keep trying. Defensive. Confluence has a different job.
• 8:39
  The attacker needs one opening. The business needs to avoid blocking the real customer standing beside that opening. It's like guarding 60 doors while other teams get unlimited players and nobody makes them go home.
• 8:51
  One o'clock expires. Cloudflare says its system retains, retrains some consistently, continuously on live traffic, creates temporary targeted rules, watches behavior updates.
• 9:03
  across different time windows, and test defensive changes before deployment.
• 9:08
  Well, if that works as described, the goal is not merely to block a known bad bot, it's to raise the attacker's cost, make everything easier. Every successful trick expire before the opponent rebuilt the playbook and before the next drive engulfs the attacker wants the same pin location every day. Cloudflare wants to move the cup between swings.
• 9:28
  And that can frustrate automation, it can also create false positives if the defense system learns badly.
• 9:35
  Automated defense needs its own guardrails.
• 9:38
  And that brings us to the deep dive. The three stories are really about control. Control loops. A control loop observes what's happening, compares it with an objective and takes an action.
• 9:49
  Checks that result and then adjusts.
• 9:51
  AI agents are becoming part of those loops.
• 9:54
  Anthropic is hardening the loop around it. On model evaluations, the British government is trying to create a loop from public problems.
• 10:01
  um, to tested product or procurement.
• 10:04
  Cloudflare is building a faster loop from a tax signal.
• 10:07
  Temporary Defense. The advantage is speed.
• 10:10
  The danger is the same speed when the measurement is wrong.
• 10:13
  Now regular businesses already have control loops even if nobody calls them that.
• 10:18
  The phone rings, somebody answers. The consumer explains the problem, we decide what happens next. The job gets scheduled, the work is completed, the invoice goes out, and the customer either returns or does not.
• 10:30
  AI can improve each of those steps.
• 10:32
  It can also quietly break that chain. So if the system answers faster but schedules the wrong technician, we improve the first step.
• 10:39
  That's how a dashboard wins while the business loses. So here's the practical demo. Deployment example that I use with my clients.
• 10:48
  Take customer intake for a contractor, plumber, real estate. team, medical officer, professional service. You give the AI one narrow role.
• 10:58
  You connect the person's name, contact information. Reason for calling, urgency, and a preferred response time.
• 11:04
  Let it draft a summary. Don't let it quote a building price.
• 11:07
  Don't let it promise an appointment that the calendar cannot support.
• 11:11
  And don't let it end. Don't let it diagnose an emergency.
• 11:13
  Route high-risk words to a human. Record every action.
• 11:17
  Measure missed calls. Response times. Completed appointments.
• 11:20
  Complaints, corrections, and human tours before and after.
• 11:24
  Then, build three boundaries.
• 11:26
  First, an authority boundary, a system that can read the approved schedule but cannot rewrite the entire calendar. Second, a money boundary. It may explain the published range, but it cannot negotiate or even Third, a communication boundary.
• 11:45
  Human Approval. Those boundaries are not a lack of confidence in AI, they're proof that we understand the job.
• 11:53
  We do not hand the apprentice every key on first morning.
• 11:57
  Well, because the apprentice gave a beautiful speech about initiative.
• 12:01
  We also need a recovery plan.
• 12:03
  If a channel's unavailable, what happens? If it sends the wrong message, how do we stop the next 9,000? If a connected account is compromised, how do we revoke access? If the consumer asks what the AI did, can we actually produce a revoke? Recovery is where promotional smoke usually leaves the room. Every demo shows the perfect past.
• 12:23
  Serious options. Operators ask what happens when the receiver slips, the ball is tipped, and somebody has to make the tackle.
• 12:30
  The risk is not only a dramatic system escape, the ordinary failures matter more because they happen more often.
• 12:36
  Hallucinated policy, misrouted customer information, an agent confidently deleting the wrong record.
• 12:43
  A security system blocking legitimate buyers.
• 12:46
  A hiring tool amplifying yesterday's bias.
• 12:49
  We should separate severity from probability.
• 12:55
  A in the attic is more likely than a meteor.
• 12:57
  We still don't give the squirrel access to the electrical panel in the company credit card. The opportunity is substantial.
• 13:04
  Small businesses can build better service without building additional costs.
• 13:07
  Workers can use AI to prepare, check, summarize, and find patterns.
• 13:13
  Families can organize decisions. Decisions that used to require hours of searching.
• 13:17
  Government can become a proving ground for useful tools and services.
• 13:21
  Instead of a museum for software contracts, security teams can respond at machine speed.
• 13:26
  But every opportunity becomes an opportunity to stronger when the authority is limited.
• 13:30
  Evidence is recorded and people can interrupt the system.
• 13:34
  Hard rails are not the brakes on deployment. They're what let us drive faster without pertaining the cliff as a feature.
• 13:41
  The move viewers can make today is simple.
• 13:44
  Choose one AI workflow already touching real information.
• 13:48
  or taking real action. Write down five things.
• 13:52
  What results should it create? What information should can it see?
• 13:56
  What actions can it take? Who can stop it?
• 13:59
  What evidence tell us that it actually works? And if any of those answers is vague, reduce the authority before expanding the intelligence.
• 14:07
  We can do that in 20 minutes. It's less exciting than asking whether superintelligence is going to love us. It can also be more useful than asking squirrels what the humans are going to do next and then building the evacuation plan.
• 14:20
  around their answer. Artificial intelligence is coming fast.
• 14:24
  We don't freeze on the shoulder. We also need to make sure don't sprint into traffic because somebody yelled disruption.
• 14:30
  We measure the speed. We understand the speed.
• 14:32
  We the distance and we choose the lane. We keep a human on that radio. The companies that win aren't going to be the ones with the bad most agents wandering through the building wearing digital badges.
• 14:42
  They're going to be the ones who know the objective, limit the authority.
• 14:45
  Watch the results and recover when realty, reality punches the brochure in the mouth.
• 14:51
  That's it. But today, the scoreboard is clear. Anthropix says a single defensive layer was not enough. The British government is demanding real,,adapt as quickly as ATT&CK.
• 15:08
  Different institutions, same operating truth.
• 15:10
  Intelligence without control. Control is not leverage, it's exposure. Our job is to use the capability, keep the data. The judgment, make the move that brings people home.
• 15:21
  AI for everyone, not just the wealthy. I'm Conor with honor.
• 15:24
  Be safe out there and I will see you tomorrow.