Latest / Dark Perimeter: Real Breaches, Real Stakes / Dark Perimeter: "The Clock Is Broken: AI, Exploits, and the Death of Monthly Patching"
Transcript
- 0:00SPEAKER_00: Welcome back to Dark Perimeter.
- 0:03SPEAKER_00: Security, AI, and the Edge of What's Coming.
- 0:07SPEAKER_00: I'm Cole Draden.
- 0:09SPEAKER_00: The twenty twenty six Verizon Data Breach Investigations report, dropped yesterday, and buried inside the data, is a finding that should stop every security leader in their tracks.
- 0:21SPEAKER_00: For the first time in 19 years of DBIR publication, vulnerability exploitation has surpassed stolen credentials as the number one breach entry point.
- 0:32SPEAKER_00: Nearly a third of all breaches now start with an attacker finding a hole in your software, not stealing a password.
- 0:41SPEAKER_00: That alone would be a headline.
- 0:44SPEAKER_00: But the reason it's happening is the story that deserves your attention tonight.
- 0:50SPEAKER_00: AI is now in the hands of threat actors, and it is compressing the window between when a vulnerability is disclosed and when it gets weaponized.
- 0:59SPEAKER_00: From months down to hours.
- 1:03SPEAKER_00: I've got two guests with me tonight who have been tracking this shift from both the technical and operational sides.
- 1:10SPEAKER_00: Dr.
- 1:11SPEAKER_00: Elliot Vance, who works at the intersection of threat intelligence and machine learning research, and Marcus Hale, who has spent years inside enterprise security operations trying to hold the line.
- 1:23SPEAKER_00: Welcome back, both of you.
- 1:26SPEAKER_01: Good to be here, Cole.
- 1:27SPEAKER_01: This DBIR finding is one of those moments where the data catches up to what practitioners have been feeling for a while.
- 1:33SPEAKER_01: The gut said something was changing.
- 1:35SPEAKER_01: Now the numbers confirm it.
- 1:37SPEAKER_02: And it's uncomfortable data because the response most organizations have to vulnerability management is still built around a monthly rhythm.
- 1:44SPEAKER_02: Patch Tuesday, the monthly vulnerability review, quarterly risk assessments, that cadence made sense when the exploit development cycle was measured in weeks or months.
- 1:53SPEAKER_02: It does not make sense anymore.
- 1:55SPEAKER_00: Let's ground this.
- 1:57SPEAKER_00: Walk me through what's actually changed on the attacker side.
- 2:01SPEAKER_00: Because we've been talking about AI-assisted attacks for a couple of years now.
- 2:06SPEAKER_00: What's different today versus 18 months ago?
- 2:10SPEAKER_01: The capability inflection point is reasoning.
- 2:13SPEAKER_01: 18 months ago, AI tools could help attackers write phishing emails or slightly speed up reconnaissance.
- 2:20SPEAKER_01: That's noise.
- 2:20SPEAKER_01: What changed is that current frontier models can reason through a vulnerability disclosure, read a CVE, read the patch diff, understand the architecture of the affected software, and begin generating working exploit code in a feedback loop.
- 2:35SPEAKER_01: The model tries something, sees why it failed, adjusts, tries again.
- 2:40SPEAKER_01: What used to require a skilled human researcher spending days on that loop now takes hours, sometimes less.
- 2:45SPEAKER_01: The DBIR specifically called out that AI is shrinking the window from months to mere hours for known vulnerabilities.
- 2:52SPEAKER_01: And the key word there is known.
- 2:54SPEAKER_01: These aren't zero days, these are CVEs that have patches available.
- 2:58SPEAKER_01: The defenders have the information they need to fix this.
- 3:01SPEAKER_01: The attackers have AI helping them race to exploit it before the defenders can act.
- 3:06SPEAKER_02: And that asymmetry is what makes this so punishing.
- 3:09SPEAKER_02: The attacker only has to exploit one gap, one system that wasn't patched in time.
- 3:13SPEAKER_02: The defender has to cover everything.
- 3:15SPEAKER_02: AI makes the attacker's side of that equation dramatically more efficient.
- 3:18SPEAKER_02: It doesn't do the same for the defender at the same rate, yet.
- 3:22SPEAKER_00: So let's talk about what this actually means if you're running a security program.
- 3:27SPEAKER_00: Because most organizations I talk to are still on a 30-day patch cycle.
- 3:32SPEAKER_00: Some are better.
- 3:34SPEAKER_00: Critical vulnerabilities in 14 days, critical infrastructure in seven.
- 3:40SPEAKER_00: But the baseline assumption in most policy documents is that monthly is acceptable.
- 3:48SPEAKER_00: Is that still true?
- 3:50SPEAKER_02: Honestly, Cole, for anything internet facing or in a high exposure zone, no, monthly is not adequate anymore.
- 3:56SPEAKER_02: And I want to be precise about what I mean, because I'm not saying you need to patch everything in hours.
- 4:01SPEAKER_02: I'm saying you need to shrink the cycle specifically for high CVSS, actively exploited vulnerabilities.
- 4:08SPEAKER_02: SISA's known exploited vulnerabilities catalog exists for exactly this reason.
- 4:14SPEAKER_02: If something hits the CAVE list, the assumption has to be that active exploitation is occurring or imminent.
- 4:19SPEAKER_02: The federal deadline for K of items is typically two weeks.
- 4:22SPEAKER_02: Private sector organizations should be treating that as a ceiling, not a target.
- 4:26SPEAKER_01: And the tooling to help defenders exists now too.
- 4:29SPEAKER_01: Vulnerability prioritization platforms can ingest threat intelligence, correlate it with your asset inventory, and surface what actually needs emergency attention versus what can wait for the normal cycle.
- 4:41SPEAKER_01: The problem is adoption.
- 4:42SPEAKER_01: A lot of organizations are still doing manual triage from a flat CSV export.
- 4:48SPEAKER_01: That approach cannot keep pace with an AI-assisted exploitation timeline.
- 4:59SPEAKER_00: Attackers use AI to compress time to exploit.
- 5:03SPEAKER_00: Defenders need AI to compress time to patch.
- 5:06SPEAKER_00: And both sides are racing on the same track.
- 5:10SPEAKER_00: Let's pivot.
- 5:11SPEAKER_00: Because the context for all of this is the state of the AI models themselves.
- 5:17SPEAKER_00: Because we can't have this conversation honestly without acknowledging where the frontier actually is right now.
- 5:24SPEAKER_00: Marcus, you've been watching the model releases this month.
- 5:28SPEAKER_00: Give us the landscape.
- 6:11SPEAKER_02: What was the Frontier model three months ago is now mid-tier.
- 6:15SPEAKER_01: The benchmark story is interesting here.
- 6:17SPEAKER_01: The traditional tests, MMLU, human evil, are essentially saturated.
- 6:22SPEAKER_01: Every frontier model scores above 90%.
- 6:25SPEAKER_01: They've lost their diagnostic signal.
- 6:27SPEAKER_01: The research community has had to move to dynamic benchmarks that test fluid reasoning, agentic task completion, and multi-step tool use in realistic environments.
- 6:41SPEAKER_01: And on those newer benchmarks, the models are still improving rapidly.
- 6:45SPEAKER_01: The one that has security implications is what the UK's AI Security Institute published recently.
- 6:53SPEAKER_01: Anthropic's Mythos Preview cleared a 32-step corporate network simulation covering reconnaissance all the way through full domain takeover.
- 7:03SPEAKER_01: It completed the simulation in three of ten runs and maintained a 73% success rate on expert-level tasks.
- 7:12SPEAKER_01: That's a penetration tester operating at a senior level, automated at machine speed.
- 7:18SPEAKER_00: That stopped me when I read it.
- 7:21SPEAKER_00: Because that's not a proof of concept anymore.
- 7:24SPEAKER_00: That's a capability that exists.
- 7:27SPEAKER_00: The question is, how long before that capability is accessible to threat actors who aren't sophisticated nation states?
- 7:35SPEAKER_01: The honest answer is that the gap between frontier capability and accessible capability has been closing faster than anyone predicted.
- 7:42SPEAKER_01: Open weight models, models where the weights are publicly released, are closing the gap with proprietary ones.
- 7:48SPEAKER_01: A model released earlier this year by a Chinese lab came in with a hallucination rate of 1.2% and a price point of 11 cents per million input tokens.
- 7:58SPEAKER_01: That's not a research artifact.
- 7:59SPEAKER_01: That's infrastructure that could be integrated into an attack platform.
- 8:04SPEAKER_00: Which brings us to the question that I think is on everyone's mind, but rarely gets asked directly on a podcast like this.
- 8:12SPEAKER_00: How close are we to AGI?
- 8:15SPEAKER_00: And I mean that with precision.
- 8:17SPEAKER_00: Not the science fiction version, but the capability threshold where AI can do essentially any cognitive task a skilled human expert can do.
- 8:28SPEAKER_00: Because that threshold matters enormously for the threat landscape.
- 8:34SPEAKER_01: The honest answer is that the people who study this most carefully have meaningfully shifted their estimates in the last six months.
- 8:41SPEAKER_01: Earlier this year, a research group tracking autonomous coding capability, basically the point at which AI can replace human software engineers at scale, moved their median estimate from late 2029 to mid-2028.
- 8:55SPEAKER_01: That's a compression of over a year in a single quarter's update.
- 8:58SPEAKER_01: Sam Altman has publicly claimed they've basically built AGI or something very close to it.
- 9:04SPEAKER_01: Satya Nadella pushed back hard and said the industry is nowhere near it.
- 9:09SPEAKER_01: Dario Amoti at Anthropic has placed his estimate between one and three years.
- 9:15SPEAKER_01: Sequoia published a piece in January, arguing that the ability to hire a frontier model today as a cognitive worker, a reasoning agent that can take actions, use tools, and complete complex tasks is itself a form of AGI that we've already crossed.
- 9:33SPEAKER_02: The definition question is real, though.
- 9:35SPEAKER_02: If AGI means a system that can beat every human at every cognitive task simultaneously, we're probably further away.
- 9:41SPEAKER_02: If AGI means a system that operates at expert level across a broad range of domains and can autonomously complete multi-step work in the real world, we are very close or already there, depending on how you define the domain boundary.
- 9:55SPEAKER_00: And from a threat perspective, the definition doesn't matter much.
- 9:59SPEAKER_00: What matters is capability.
- 10:02SPEAKER_00: The question isn't whether we cross some philosophical threshold.
- 10:07SPEAKER_00: The question is whether adversaries have access to systems that can reason, plan, adapt, and execute at expert level with minimal human involvement.
- 10:20SPEAKER_00: That question has a current answer increasingly yes.
- 10:26SPEAKER_01: The agenic piece is the key development to watch.
- 10:29SPEAKER_01: We've gone from models that answer questions to models that reason to models that now plan multi-step actions, use tools, call other AI systems, and execute tasks autonomously.
- 10:40SPEAKER_01: That last step is qualitatively different.
- 10:45SPEAKER_01: Find a vulnerability in this organization's perimeter, develop an exploit, establish persistence, and work toward it without requiring a human in the loop at each step.
- 10:55SPEAKER_01: The time horizon for autonomous action is expanding every quarter.
- 11:00SPEAKER_00: So, where does that leave a security practitioner sitting at their desk tonight?
- 11:06SPEAKER_00: Let's bring this down to earth.
- 11:08SPEAKER_02: A few concrete things.
- 11:09SPEAKER_02: First, your patch management policy needs a tiered review now, not next quarter.
- 11:13SPEAKER_02: Build an emergency lane that bypasses the monthly cycle for anything that hits the SISAKEV list or scores critical with active exploitation evidence.
- 11:20SPEAKER_02: That lane needs to have teeth.
- 11:23SPEAKER_02: Uh defuncy, defined SLA, defined escalation, and the authority to act without waiting for the next change management window.
- 11:31SPEAKER_02: Second, if your vulnerability triage is still manual, you need to automate prioritization.
- 11:36SPEAKER_02: Not because automation is a trend, but because the attacker's timeline now moves faster than a human analyst can read and respond to a flat list of CVEs.
- 11:49SPEAKER_02: Third, the DBIR finding about mobile social engineering deserves attention too.
- 11:54SPEAKER_02: As people get more sophisticated about email phishing, attackers are moving to voice calls and text messages with a success rate 40% higher than traditional email phishing.
- 12:06SPEAKER_02: Your user awareness program needs to address that shift specifically.
- 12:10SPEAKER_01: And at the strategic level, if you haven't started thinking about your cryptographic inventory, now is the time.
- 12:16SPEAKER_01: The quantum timeline is compressing.
- 12:18SPEAKER_01: Research published this year suggests breaking RSA 2048 may require far fewer qubits than we thought two years ago.
- 12:25SPEAKER_01: NIST has set this year as the migration deadline for quantum-resistant algorithms.
- 12:32SPEAKER_01: The harvest now decrypt later threat is not theoretical.
- 12:37SPEAKER_01: Adversaries are storing encrypted data today, betting they can crack it in a few years.
- 12:42SPEAKER_01: If your organization handles data that needs to remain confidential for more than five years, that is an active risk right now.
- 12:55SPEAKER_00: AI has changed the speed of offense.
- 12:58SPEAKER_00: It's in the process of changing the speed of defense.
- 13:02SPEAKER_00: The organizations that fall behind on adopting AI-assisted defensive capabilities while adversaries are already running AI-assisted offensive capabilities are going to find that gap very hard to close.
- 13:15SPEAKER_00: Monthly patch cycles assumed a world where attackers moved slowly.
- 13:20SPEAKER_00: That world is gone.
- 13:36SPEAKER_00: Well said.
- 13:38SPEAKER_00: Elliot Marcus, as always, appreciate you both coming in and thinking through this with clarity.
- 13:45SPEAKER_00: We'll have links and context from tonight's episode at the usual place.
- 13:50SPEAKER_00: Stay sharp, stay current, and we'll see you on the next one.
- 13:55SPEAKER_00: This is Dark Perimeter.