Every Day Is Zero Day | Wiz CTO Ami Luttwak

Original Media · 46m · English · Podcast · The a16z Show

TLDR

AI has fundamentally changed cybersecurity: models now autonomously find and exploit complex vulnerabilities in ~20 minutes, while open source models have caught up to frontier capabilities in just 3 months. The patching ecosystem is broken — exploits arrive before patches, and distribution takes months. Defenders must adopt continuous automated red teaming, focus exclusively on critical risk over compliance, and design systems assuming vulnerabilities will exist. We are returning to a 1990s-era of constant breaches where AI attacks surpass human capabilities, and only companies building security automation and secure design from day one will survive. Public infrastructure with legacy software is critically exposed. The world is not ready for machine-speed attacks. Every day is now zero day.

For you

PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.

Sign in to unlock

Surprising Insights

•

Open source models are only ~3 months behind frontier models in offensive capabilities, not the 6-12 months previously expected.

•

The PHP vulnerability (CVE-2024-4577) is described as 'probably the most severe vulnerability ever found by AI.'

•

915 AI development platforms were attacked in October alone, far exceeding the initial report of 9 companies.

•

Time to patch has actually increased over the last 20-25 years despite automation advances, while exploit development time has collapsed.

•

AI is now better than humans at complex vulnerability exploitation — the only domain where this is currently true.

•

Worms may return like the Windows XP era because AI-driven infection speed now exceeds patch deployment speed.

•

Information about vulnerabilities frequently leaks before CVEs are published and before patches are available, leaving defenders with zero response time.

•

A single hard-coded key in a mobile application recently exposed all user data, demonstrating that unauthenticated surfaces remain the primary attack vector, not complex code vulnerabilities.

Key takeaways

•

AI models can now autonomously find and exploit complex vulnerabilities, collapsing attack timelines from weeks/months to approximately 20 minutes from initial access to full compromise including secret exfiltration.

•

Open source models are catching up to frontier models in offensive capabilities in roughly 3 months rather than the expected 6-12 months, democratizing nation-state-grade attack capabilities to anyone with compute access.

•

The patching ecosystem is fundamentally broken: exploits now arrive before CVEs are published and before patches exist; even when patches exist, distribution through Linux distros takes weeks to months, while AI exploits execute in minutes.

•

Continuous automated red teaming (attacking your own systems with the latest AI models) is now the only viable defense because 'whatever was secure last week may not be secure today' as model capabilities improve weekly.

•

Security teams must stop spending most of their time on compliance, reporting, and configuration checks (CSPM-style) and focus exclusively on finding critical risk, mitigation, detection engineering, and response readiness.

•

New companies should design fully automated patch pipelines from day one — agentic development, testing, CI/CD, and DevOps — because legacy companies cannot realistically retrofit this across thousands of developers and obscure ownership.

•

AI attacks have surpassed human attackers in complex zero-day exploitation (e.g., kernel driver zero-days used daily), making this the only domain where AI currently outperforms humans, and this asymmetry favors attackers.

•

Public infrastructure (refrigeration, traffic systems, food supply chains) running legacy software connected to the internet is critically exposed with minimal cyber budgets; Scan for Good provides initial scanning to push these entities toward self-sustaining red teaming.

Notable quotes

“We are calling this Black October since it's, in my view, like a Black Swan event.”

▸ Notable quotes · 2:11

“We've never seen, never, ever, seen so many security incidents happen in a single month.”

▸ Notable quotes · 2:15

“Models are good enough now to find and exploit complex vulnerabilities.”

▸ Notable quotes · 4:23

“The other is that the open source models, right, that's the scary bit.”

▸ Notable quotes · 4:36

“They're able to find ways to run code on your environment... run an entire attack that usually took for the best attackers weeks or months in, again, around 20 minutes.”

▸ Notable quotes · 4:53

“The time to get an exploit dramatically went down... Many times the information leaks before the CV even is out.”

▸ Notable quotes · 8:50

Chapters

•

0:00 — Black October Introduction: Ami Ludwak describes October 2024 as a 'Black Swan' event with unprecedented volume and severity of security incidents, including the PHP vulnerability found by AI and a surge in AI-driven attacks against companies and infrastructure at unprecedented scale and speed.

•

2:11 — AI Offensive Capabilities & Open Source Models: Two key shifts: models are now capable of finding and exploiting complex vulnerabilities autonomously, and open source models have caught up to frontier models in just ~3 months, democratizing offensive capabilities. Attackers can now go from initial access to full compromise in ~20 minutes.

•

5:20 — PHP Vulnerability Case Study: The PHP vulnerability (CVE-2024-4577) affected ~70% of the internet and nearly 100% of Wiz customers. While 50% patched within 2-3 days, full resolution will take years; many systems have no patch path. This illustrates the broken patching ecosystem.

•

8:50 — Patching Challenges & AI-Assisted Patching: Exploit development time has collapsed while patch distribution has slowed. Information leaks before CVEs; Linux kernel patches take weeks to reach distros. Automated patching requires full agentic pipelines (dev, test, CI/CD, DevOps) — feasible for new startups, unrealistic for legacy companies.

•

10:50 — Vulnerability Research with AI / Atlas System: Wiz built 'Atlas,' a harness for vulnerability detection using multiple frontier models (OpenAI, Google, Anthropic). Each model has unique capabilities; multiple teams now independently discover the same vulnerabilities simultaneously as new models unlock new vulnerability classes.

•

13:40 — Multi-Model Pipelines & Cost Optimization: Effective AI security requires intelligent pipelines: deterministic detection → cheap model → advanced model → most advanced model for final verdict. Cost management across model tiers is the core engineering challenge for scalable AI security products.

•

16:20 — Continuous Testing Paradigm Shift: Static scans are obsolete. Continuous, budget-aware scanning across code, runtime, external exposure, and active attacks is required. Scanning external unauthenticated attack surface is more urgent than code scanning. Intelligent budget allocation across scan types is the new challenge.

•

18:50 — Cloud Security & AI Integration: Cloud security fundamentals remain — fix critical risk — but AI adds new vulnerability classes. Customers now expect AI-native automation: not just vulnerabilities but auto-generated fixes, PRs, full investigations, forensics, root cause analysis, and automated response.

Speakers

•

Sarah Wang — Host / Investor

•

Zane Lackey — Host / Investor

•

Ami Ludwak — Guest / CTO

⋮