By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Cyberattacks Used to Take Months. AI Can Do It in Minutes.

AI isn’t just helping defenders find vulnerabilities. It’s giving attackers the speed and expertise to exploit them faster than many companies can respond.

Zane LackeyhostAmi LuttwakguestSarah Wanghost

Nov 13, 202646mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

Great conversations run long. Their best ideas don't have to.

uListen is free on iPhone, iPad, Android, Chrome, and Claude.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

Why security is “going back to the ’90s”: constant breaches and AI-speed attacks
The conversation opens with the claim that the security landscape is shifting back toward an era of frequent, high-impact breaches. The speakers frame AI as changing both attacker capabilities and the defender’s need to automate across offense and defense.
- •Zane Lackey says security will resemble the ’90s/2000s again, with breaches “popping out all the time.”
- •Ami Luttwak describes attackers using AI at unprecedented scale and speed against companies and startups.
- •Sarah Wang references the idea that “every day is zero day” and asks what AI is doing to cause this.
- •The group frames automation on both “blue” (defense) and “red” (offense/testing) as increasingly necessary.
“Black October”: a surge of incidents, vulnerabilities, and AI-driven exploitation
Ami Luttwak explains why security professionals are calling the period “Black October,” describing it as an unusually intense month for both the number and severity of incidents. He ties this to AI models becoming capable enough to discover and exploit complex vulnerabilities, with open-source models rapidly catching up.
- •Sarah Wang asks for an explanation of what’s happening in “Black October” and how to get through it.
- •Ami Luttwak calls it a black swan-like event with an unprecedented number of incidents in a single month.
- •He highlights both increased volume and increased severity of vulnerabilities discovered.
- •He emphasizes that the rise of AI attacks is “even worse,” due to new speed and expertise.
From warnings to reality: models, open source catch-up, and 20-minute end-to-end attacks
The speakers trace the path from earlier warning moments to the current wave of practical attacks. Ami Luttwak argues that open-source model catch-up means attackers can find and operationalize flaws quickly, turning API mistakes or other issues into full compromises in minutes.
- •Ami Luttwak cites two earlier moments: the “Mythos” moment and the “HackingFace incident with OpenAI,” which didn’t immediately change day-to-day reality.
- •He says “Black October” is when these theories “crashed upon us” and became immediate and widespread.
- •He argues models can now find and exploit complex vulnerabilities, and open-source models are catching up faster than expected.
- •He claims attackers can run an entire attack that used to take weeks or months in about 20 minutes, including leveraging simple API mistakes.
- •Ami Luttwak notes AI development platforms are being targeted for AI keys, and the list of attacked companies is growing.
PHP2Shell case study: prevalence, rapid patching—then a long tail of exposure
Using the PHP vulnerability as a case study, the speakers discuss how widespread exposure can be even when patching begins quickly. Ami Luttwak stresses that ecosystem realities—legacy versions and lack of patch availability—mean vulnerabilities can linger for years.
- •Zane Lackey asks about prevalence across customers and patching progress weeks later.
- •Ami Luttwak describes PHP as extremely prevalent externally, affecting nearly all customers in some way.
- •He reports about 50% patching within two to three days, driven by heightened awareness.
- •He notes that because PHP is so widespread, resolving the vulnerability across the internet could take years.
- •He highlights that some environments cannot patch at all, and older PHP versions may not receive patches.
Why patching can’t keep up: leaks, delayed patches, and the limits of auto-patching
The discussion turns to the widening gap between exploit development speed and patch availability or adoption. Ami Luttwak describes how details may leak before a CVE or patch is available, and argues that fully automated patching is difficult for legacy organizations—though startups can design for it upfront.
- •Sarah Wang contrasts faster exploit creation with patch times that haven’t improved and asks whether AI can help patching.
- •Ami Luttwak says exploit timelines have “dramatically” decreased and may be worse when information leaks early.
- •He notes patches may lag CVEs, and distro pipelines (like kernel distributions) can take weeks or months to propagate fixes.
- •He argues the industry isn’t ready for AI exploitation “in minutes.”
- •He’s skeptical about a broad path to automatic patching for legacy companies, but says startups can design for auto-patching with automated dev, testing, CI/CD, and DevOps.
How Wiz uses AI for vulnerability research: multiple models, harnesses, and “Bitcoin mining” dynamics
Zane Lackey asks how AI is being leveraged in vulnerability research, and Ami Luttwak describes an environment where new model releases unlock new classes of findings. He explains that Wiz uses an internal harness and rotates across models, expecting more simultaneous discoveries across teams.
- •Zane Lackey references independent discovery by Wiz and another group and asks how AI enabled the research process.
- •Ami Luttwak predicts this will happen more often as new models bring new capabilities.
- •He compares vulnerability discovery to “Bitcoin mining,” where new tools uncover more opportunities.
- •He describes using an internal harness called Atlas with different models because each has different strengths.
- •He argues many teams will find vulnerabilities at the same time, especially around critical infrastructure targets.
Model orchestration as the real engineering challenge: balancing capability, speed, and cost
Sarah Wang probes how different models perform and how independent cyber companies work with frontier labs. Ami Luttwak emphasizes that no single model is best; the key is building efficient pipelines that route tasks across models to manage cost and scalability.
- •Sarah Wang asks what different models are better at and how to think about relationships between frontier labs and cyber companies.
- •Ami Luttwak says model selection is a “complicated matrix,” with frontier and cyber-tuned variants behaving differently.
- •He highlights tradeoffs across model behavior, speed, and cost, with no universal answer.
- •He argues scalable services require intelligent pipelines that mix models for different task complexity.
- •He describes a “Blue Agent” pipeline that escalates from deterministic detection to cheaper models to the most advanced model before alerting customers.
The hidden bill of continuous security testing: why naive scanning explodes budgets
The speakers discuss continuous testing and why point-in-time scanning approaches can be financially unsustainable. Ami Luttwak gives examples of expensive code scanning and argues for continuous, intelligent scanning that uses budget effectively across multiple techniques.
- •Sarah Wang asks about the shift from static to continuous testing and whether costs can come down.
- •Ami Luttwak says scanning all code with a new capability can be extraordinarily expensive for large codebases.
- •He argues repeating scans for each new model/version is not viable as capabilities and code constantly change.
- •He notes some customers spend tens of millions on scanning and says this doesn’t create a sustainable continuous posture.
- •He frames the challenge as allocating budget intelligently across inside/outside scanning, active/passive approaches, code and runtime analysis.
Cloud security in the AI era: the basics still matter, but expectations shift to agentic workflows
Zane Lackey asks about the state of cloud security and how AI changes it. Ami Luttwak says attackers still begin with familiar mistakes, but AI pushes customers to demand automation across remediation, investigation, forensics, and eventually response.
- •Zane Lackey asks first to set AI aside and describe cloud security state, then layer AI in.
- •Ami Luttwak says AI attackers still start from classic vulnerabilities and misconfigurations; old problems don’t go away.
- •He says organizations need a system of record for visibility and to focus on critical risk and fixing it.
- •He notes customers now expect more automation: finding code, creating fixes, and generating pull requests.
- •He says detections are no longer enough; customers want full investigations, forensics, root cause, and help pushing attackers out, raising the prospect of automated response.
Red Agent: “attack yourself” continuously as defenses become dynamic
Zane Lackey asks about Red Agent and whether Wiz is giving customers machine-speed defense. Ami Luttwak argues defenses must become dynamic because every new model increases attacker capability, making continuous self-attack with the latest models essential to staying ahead.
- •Zane Lackey asks for details on Red Agent and the goal of delivering machine-speed defense.
- •Ami Luttwak frames defense as two forces: blue to detect/stop attacks and red to attack yourself first.
- •He says static defenses no longer work because new models create new attacker capabilities over time.
- •He uses an analogy of hiding a key under a mat becoming insecure as attackers gain “superpowers.”
- •He describes Red Agent as running offensive AI from the outside using inside context to find issues before attackers do, emphasizing speed.
Agent swarms and layered red-teaming: prioritize external exposure before deeper code work
The conversation dives into how red-team automation is structured and budgeted. Ami Luttwak describes a layered pipeline: deterministic exposure analysis first, then unauthenticated external attack-surface scanning, then deeper pen-testing and code analysis with varying time and model investment.
- •Zane Lackey asks whether red-side pipelines also layer deterministic methods and cheaper models before expensive ones.
- •Ami Luttwak says they use “everything,” emphasizing budget-efficient layering rather than brute-force scanning.
- •He says the most urgent step is deterministic analysis of exposure and unauthenticated internet-facing surfaces.
- •He describes scaling: scanning millions of endpoints quickly, while allowing long-running efforts for critical apps.
- •He ties red and blue together: if a fix is too expensive, route to blue-side detections and readiness instead.
What founders must do before shipping: parallelize security and prepare for fast theft of secrets
Sarah Wang asks what founders should do if they have a release pending and security backlog. Zane Lackey advocates parallelizing release and security through automation and continuous testing, while Ami Luttwak describes the lived reality of rapid AI-driven compromises that steal secrets and keys quickly.
- •Sarah Wang asks how a founder should decide what must happen before shipping when security issues exist.
- •Zane Lackey says the security game has changed and emphasizes automation on both blue and red sides via agents.
- •He recommends doing work in parallel—continuous testing and building patching automation from the beginning for new companies.
- •Ami Luttwak describes startups being attacked, discovering incidents after delays, and facing rapid theft of secrets and customer AI keys.
- •He notes startups are targets, especially if they have valuable assets like AI keys, and mentions limiting keys to reduce damage from theft.
Where security spending should (and shouldn’t) go: tools alone won’t solve it
Sarah Wang asks where budgets should increase and decrease. Ami Luttwak argues the core issue isn’t simply higher budgets or buying tools; it’s security architecture and design built for elite-level attackers, plus continuous scanning and isolation strategies.
- •Sarah Wang cites a CISO expecting a large budget increase and asks where money should go and what becomes obsolete.
- •Ami Luttwak says budgets may rise but won’t solve the problem alone; focus and architecture matter more than purchasing tools.
- •He argues many “latest security tools” are no longer relevant, even while AI-native tools are needed.
- •He emphasizes secure design elements: sandboxes, isolation, network and process lockdown, image lockdown, and designing for inevitable vulnerabilities.
- •He warns against assuming platforms are inherently safe, noting zero days can exist across layers like kernels and container boundaries.
Most security work isn’t security: drop report-driven work and focus on real risk and readiness
The episode closes by challenging how security teams spend their time and what “security” actually means under AI pressure. Ami Luttwak argues teams must cut non-essential work, prioritize attacker-focused risk, and build sustainable isolation strategies—because response time needs to shrink dramatically.
- •Ami Luttwak says most of what security teams do is not security and suggests measuring time spent on red teaming versus compliance/reporting.
- •He defines real security as attacking yourself, building guardrails, and preparing for attacker arrival.
- •He calls for efficiency: focus on finding critical risk, mitigating it, adding detections, and response readiness.
- •He criticizes “classic CSPM” as configuration-check lists divorced from attacker paths, advocating prioritization based on how attackers would enter.
- •He notes that legacy complexity (ownership, context, process) makes rapid automation hard, but argues relevance depends on responding in minutes.
The next few years: Black October as the beginning, and “take it off the internet” if it’s not ready
Zane Lackey predicts a return to frequent unauthenticated zero days and even the reemergence of worms, comparing patching challenges to past eras. Ami Luttwak says open-source models are catching up faster than expected, that AI may outperform humans in this domain, and that survival requires redesigning what’s exposed to the internet.
- •Zane Lackey predicts frequent unauthenticated zero days and more constant breaches, with a possible return of internet worms.
- •He shares an analogy from past operating system patching where systems could be infected faster than they could be updated.
- •Ami Luttwak says Black October may be only the beginning and that open-source model catch-up is happening faster than hoped.
- •He claims AI attacks are prevalent, easy, and better than any human, with complex zero-day attacks occurring very frequently.
- •He argues companies must redesign internet-exposed systems for this reality, and anything not ready should be taken off the internet quickly.
Protecting public infrastructure: “Scan for Good” and the risk to under-resourced services
Sarah Wang shifts the conversation from enterprises to public infrastructure and organizations without large cyber budgets. Ami Luttwak explains why he’s worried about any internet-connected legacy systems and describes Scan for Good as an effort to help public services scan for problems and catalyze ongoing self-scanning.
- •Sarah Wang asks how worried we should be about attacks on public infrastructure and entities lacking cyber budgets.
- •She references incidents involving refrigeration units and raises concerns about broader consequences.
- •Ami Luttwak says he is worried and that anything connected to the internet—including legacy real-world systems—can be at risk.
- •He explains Scan for Good aims to help public services by scanning and finding issues, even if it can’t solve everything.
- •He hopes findings encourage organizations to treat it as a larger project and begin ongoing red scanning themselves.

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.