How to Defend Against AI Hackers, with Wiz’s Ami Luttwak
Summary
- Ami Luttwak says AI can now complete attacks that once took elite hackers weeks or months in roughly 20 minutes. Open-source models are catching up faster than he expected; he puts the frontier advantage in vulnerability discovery at “not more than three months.” The immediate targets include startups holding customers’ AI keys, where stolen credentials create an opportunity for “LLM jacking.”
- Faster exploitation is colliding with patching systems that cannot move at machine speed. Luttwak reports roughly 50% patching within two or three days for the PHP vulnerability across Wiz’s customer base, but older versions lack patches and distribution delays can last weeks or months. He sees an automation path for new companies with automated development, testing, CI/CD, and operations; legacy organizations face a much harder transition.
- Security AI faces a substantial cost problem alongside its capability gains. A comprehensive code scan might cost $5 million, and Luttwak says some customers are spending tens of millions on scanning. Sustainable services require deterministic checks and cheaper models to filter work before expensive models investigate: otherwise, “you’re gonna lose money on your sales.”
- Luttwak puts continuous testing of external exposure ahead of another comprehensive code scan. Newly capable models can make yesterday’s defenses inadequate, so testing must recur as capabilities change. He says a majority of Wiz customers already use its red agent, while broader industry adoption remains insufficient: “The best defense is attack.”
- Founders holding valuable credentials need security investment and automation earlier in their development. Luttwak says a list of AI development providers attacked in October grew from nine to 15 companies. He recommends limits on AI keys to contain stolen-key charges; Zane Lackey argues for continuous testing and patching alongside development when founders face release deadlines.
- Luttwak argues that larger budgets cannot substitute for secure architecture. Sarah Wang cites one CISO expecting an 80% budget increase next year, but Luttwak emphasizes isolation, lockdowns, and preparing for nation-state-level attackers. His challenge to checklist products and reporting-heavy workflows is blunt: “Most of what security teams do is not security.”
- The forecast extends beyond software companies, with substantial uncertainty about the next two years. Lackey says he would not be surprised to see internet worms return; Luttwak says October “might be the beginning” but cannot foresee two years ahead. Refrigeration attacks and other connected infrastructure motivate Wiz’s Scan for Good initiative, which offers public services an initial scan rather than a complete solution.
Deep dive
1. Black October compresses the attack cycle to 20 minutes
Luttwak calls October a “black swan event”: unprecedented numbers and severity of vulnerabilities coincided with actual AI attacks. Before Black October, the Mythos moment showed that models could find vulnerabilities, while the Hugging Face incident involving OpenAI showed the risk of AI attacks, though those attacks were accidental. Now models can find and exploit complex flaws. An attack that took the best human attackers weeks or months can run in “around 20 minutes,” including exploitation of an ordinary API mistake.
Lackey relays that offensive-security practitioners have described open-source models as six to nine months behind. Luttwak puts the advantage in vulnerability discovery at “not more than three months,” while distinguishing discovery from executing attacks. Open-source models are already good enough for the latter. His previous hope that they were a year behind has proved too optimistic.
Startups are already targets because their credentials have value. Luttwak says Wiz’s list of AI development providers attacked in October expanded from nine to 15 companies and kept growing. Attackers can steal customers’ AI keys before a startup detects and responds. His practical recommendation: put limits on those keys so theft does not produce enormous usage charges.
Independent discovery is becoming more common: Wiz and another team, Caliph, found PHP to Shell separately. Luttwak compares vulnerability discovery to “Bitcoin mining”: new models unlock additional findings. Wiz’s zero-day competition is also receiving multiple reports of another critical vulnerability he believes will significantly affect the internet. The defensive objective is to use frontier models’ brief advantage on critical infrastructure before attackers arrive.
2. Patching remains constrained by the systems around it
Luttwak describes the PHP flaw as “probably the most severe vulnerability ever found by AI.” Almost 100% of Wiz customers ran some PHP; he reports 50% patching within the first two or three days. Sarah Wang estimates that PHP underpins roughly 70% of the internet, and Luttwak says this single vulnerability could take years to resolve. Older versions are not even receiving patches.
Wang’s pushback is the central asymmetry: exploit-writing time has fallen, while patching time appears to have increased over the last 20–25 years. Luttwak says information can leak before the CVE is out, and even after a CVE is out, a patch may not be ready. Wiz worked with PHP for months to prepare the industry, but information leaked early. Customers then had no patch to apply, however quickly they wanted to respond.
Distribution adds another delay outside a customer’s control. A reported Linux kernel vulnerability might take weeks or months to reach every distribution, while AI can exploit flaws in minutes. “The world isn’t ready for this,” Luttwak says. Asked about automatic patching, his answer is candid: “We don’t yet know exactly how to patch automatically,” and “I honestly don’t see a path toward that.”
His qualification matters: a new company can design for automatic patching from inception. That requires an entire pipeline—an agent turns a ticket into a patch, automated tests validate it, and automated CI/CD and DevOps deploy it. Most legacy companies lack that chain, and he does not consider it realistic for much of their existing code. For startups, building it should be an explicit task.
3. Model orchestration determines whether security AI can scale
Wiz’s Atlas harness uses different models for vulnerability detection because capabilities differ across providers, versions, and specialized cyber variants. Luttwak describes a “complicated matrix” of capability, speed, and cost. Even with Wiz now part of Google, his prescription for an independent company is to “try all the models all the time” and find the best combination.
The blue agent illustrates the economics. Deterministic detection comes first, followed by a cheap model and then a more advanced model. If those stages judge an event malicious, the most advanced model makes the final assessment before notifying the customer. Running the costliest model on every event would not work. This pipeline design is, for Luttwak, a central engineering challenge for scalable AI services.
Continuous scanning compounds the expense. A hypothetical April code sweep might have wasted $5 million; then “Mythos 1.2” arrives—do you repeat it? Some customers are spending tens of millions. Budgets must span external and internal testing, code, disks, and runtime as models and applications change. Luttwak explicitly declines to claim victory: “I’m not going to tell you that we’ve solved it.”
4. Continuous external testing takes priority over blanket code scanning
The red agent’s premise is that security changes even when an application does not. Luttwak’s analogy is a door with a key hidden under the mat: attackers acquire “superpowers every week,” so what appeared secure last week may fail today. Continuous attacks with the latest models aim to identify that change before an adversary does, using internal context to guide external testing.
His ordering is emphatic: deterministically map exposure, then test everything unauthenticated and reachable from the internet before deeper penetration testing and code analysis. “Who cares about your code?” is his provocation about urgency. The concrete example is a recent incident involving a key hard-coded in a mobile application that exposed people’s data. Attackers exploit accessible mistakes, not just sophisticated software flaws.
Effort should vary with the target. An exposure agent might spend a few minutes per endpoint across a million endpoints; a critical application might warrant 10 hours of attempted exploitation. Red and blue workflows should also connect: if engineering says a discovered problem is too expensive to fix, the blue agent can place detections around it. “Everything is going to be a pipeline now.”
A majority of Wiz customers already use the red agent, which Luttwak calls one of its fastest-growing features; he says coverage is moving quickly toward almost complete adoption. Across the industry, he sees insufficient uptake and misplaced emphasis on code-scanning publicity. “There’s nothing more important than continuously scanning your environment from the outside.”
5. Security teams need architecture, context, and response at machine speed
Wang asks how founders with security backlogs should approach an upcoming release, citing a note from Lackey and colleague Joel suggesting more security spending and potentially delayed launches. Lackey’s answer emphasizes doing the work in parallel: agents continuously test while development proceeds, and new companies build patching automation into software and infrastructure development from the beginning. “The security game has just changed.”
Wang’s example of an 80% budget increase comes from one CISO’s expectation for next year. Luttwak questions whether budgets alone can solve the problem: companies must prepare for attackers with elite expertise and speed. His priorities include sandboxes, isolation, and network, process, and image lockdowns. Simply placing workloads in Kubernetes is insufficient when kernel vulnerabilities permit container escapes. “Get ready for a nation-state attacker.”
His critique of existing workflows is deliberately sharp: “Most of what security teams do is not security.” He contrasts time spent attacking systems with compliance and report production, and singles out classic CSPM configuration-check lists. The work he prioritizes is finding critical risk, mitigating it or adding detections, preparing responses, and building a sustainable isolation strategy.
Customers increasingly expect fixes and pull requests, plus investigations, forensics, and root-cause analysis. Automated response is the next step, which Luttwak calls “very scary.” Legacy organizations still struggle with thousands of developers, unclear resource ownership, and unwritten knowledge; an agent does not instantly supply that context. The operational test remains unforgiving: “If I cannot respond in minutes,” he asks, “what am I relevant for?”
6. The next breach wave could reach connected physical infrastructure
Looking two to three years ahead, Lackey predicts a return to the 1990s and 2000s: frequent unauthenticated remote vulnerabilities and breaches. He “wouldn’t be surprised” to see worms again. His Windows XP analogy is specific: internet-connected machines could become infected during installation and updating, forcing users to install offline because compromise happened faster than patching.
Wang recalls Luttwak’s April warning of a Y2K moment in 12–18 months. He now says October “might be the beginning,” while declining to foresee two years ahead. His categorical claim is that AI attacks are already “better than any human’s”; he cites that day’s kernel-driver zero-day attack. He thinks survival requires redesigning internet-facing systems for these attackers and quickly taking unready systems offline.
Wang brings the discussion to European refrigeration attacks and food-supply consequences. Luttwak worries about any connected legacy system, including traffic infrastructure. Scan for Good offers public services scans to identify problems and encourage sustained testing. He limits the promise: “It’s not going to solve all their issues.” The hope is that concrete findings demonstrate why a larger security project is necessary.