Resilient Cyber

Zack Korman on why the AI safety debate has been captured by an extinction narrative, and why the incident everyone is calling a watershed moment looks a lot more like a preventable security failure.

In this episode I sit down with Zack Korman, co-founder of Embroidery, where he uses AI to monitor AI agents, and former CTO of the cybersecurity company Pistachio. I first came across Zack on X during the controversy over fake SOC 2 reports, and he has since become one of the sharpest critics of how AI safety is being framed, funded, and investigated.

We get into the effective altruism roots of the existential risk movement, why he argues METR's review of the Hugging Face incident was not independent oversight, and what an actual incident response firm would have done differently. Zack makes the case that alignment is one control among many, that there is no "safe" path in AI, only trade-offs, and that the real risk for most organizations is enterprise environments that were never ready for agents in the first place.

In this episode:

● How building an AI insider threat product pulled a developer and CTO into the cybersecurity community, and into picking fights on X
● Effective altruism, longtermism, and how a focus on preventing extinction came to dominate AI safety
● Why "just be careful" misses the point when every path involves trading one risk for another
● Dario Amodei's Pacing the Frontier, independent auditors, and why Zack calls bringing in METR "bring your friend to work day"
● The funding and relationships connecting METR, Redwood Research, Coefficient Giving, and the labs
● True believers versus IPO hype, and why genuine belief does not make someone right
● Why many AI doomers believed we were all going to die before they ever learned about computers
● The Hugging Face investigation, context drop in AI-analyzed transcripts, and treating knowable facts as unknowable
● What Unit 42 or Mandiant would have demanded before putting their name on the report
● Eight layers of failure, from a single package proxy at egress to missing monitoring, classifiers, and kill switches
● Alignment failure versus containment failure, and why alignment belongs inside defense in depth
● The real risks: threat actors misusing AI, enterprise agent deployments, and new attack chains
● Why security's risk-averse, laggard culture may be its biggest vulnerability

Chapters:

0:00 Intro
0:54 Zack's Background
1:43 What Turned a CTO Into an AI Safety Critic
3:31 Effective Altruism and the Extinction Narrative
5:53 No Safe Path, Only Trade-offs
6:40 METR, Redwood, and Bring Your Friend to Work Day
10:23 True Believers, Hype, or Both
12:49 How Doomers Find Their Way Into AI
14:39 Taking AI Risk Seriously When It Is Ideological
16:56 What an IR Firm Would Have Asked
21:10 Eight Layers of Failure in the Hugging Face Incident
22:49 Alignment Failure Versus Containment Failure
24:32 Alignment as One Layer of Defense in Depth
26:22 Why Enterprise Environments Are Not Ready for Agents
29:10 Security's Laggard Culture
29:19 Closing

Connect with Zack Korman:

X: https://x.com/ZackKorman
Embroidery: https://embroidery.io
Website: https://zkorman.com

Resilient Cyber: https://www.resilientcyber.io

Subscribe for more conversations with security practitioners and leaders.

#aisafety #aisecurity #agenticai #incidentresponse #effectivealtruism #cybersecurity

Jonathan Rende of Checkmarx on why the model writing your code cannot also be the control that validates it, and why rules-based and AI-driven scanning turn out to find almost entirely different bugs.

In this episode I sit down with Jonathan Rende of Checkmarx. Jonathan worked with Fortify and SPI Dynamics back in the day, spent most of the last decade leading product teams in developer and DevOps tooling, and came back to security eighteen months ago because, as he puts it, this is the heart of the hurricane. His argument is that AI is a bigger disruption than the internet, SaaS, or mobile were, not because of any single capability, but because it hits roles, process, and productivity all at once.

We get into the two waves he has watched play out with CISOs and their CEOs, why the pendulum has swung back toward program and posture questions in the last quarter, what his research team found when they benchmarked deterministic and probabilistic scanning side by side, and why he thinks agentic AppSec raises the profile of the security team rather than automating it away.

In this episode:

● Why the first half of 2026 became a real inflection point rather than another AI talking point

● The two waves: engineering told to run at any cost, then the pendulum swinging back toward posture and program design

● Why functional AI-generated code and secure AI-generated code are still two different things

● Separation of church and state, and the conflict of interest in letting the model that generates code also validate it

● Benchmarking deterministic and AI-based scanning across dozens of open source projects, and why the overlap stayed consistently under 10%

● Fidelity, F1 scores, and the absence of real standards or shared benchmarks in AppSec

● Why an incentive to reduce risk and an incentive to sell tokens are not the same incentive

● Why agents free AppSec professionals for higher-order work, and why this is not a dark factory

● Shadow IT becoming shadow AI, and early scans where half surfaced models, agents, and MCP servers security teams did not know existed

● Why new threat vectors show up first in fast-moving unregulated companies while regulated ones see more code-level issues

● Low-priority vulnerabilities chained into real impact, and why backlogs now matter as much as incoming code

● What to change first: metrics defined up front, in-workflow AppSec, and security reviews that went from annual to monthly

Chapters:

0:00 Intro
0:18 Fortify, SPI Dynamics, and a decade in developer tooling
1:24 Why he came back to AppSec
2:54 Why the first half of 2026 was the inflection point
5:27 Two waves, and the pendulum swinging back
9:08 Functional AI code versus secure AI code
11:58 Layered defense and the condensed lifecycle
15:04 What agents free AppSec teams to actually do
17:50 Separation of church and state
20:15 Fidelity, F1 scores, and not selling tokens
23:25 New threat vectors and organizational maturity
25:47 Shadow AI and what the inventory scans found
27:58 What to change first in your AppSec program
30:44 Closing

Connect with Jonathan:

LinkedIn: https://www.linkedin.com/in/jonathanrende/

Checkmarx: https://checkmarx.com

Resilient Cyber: https://www.resilientcyber.io

Subscribe for more conversations with security practitioners and leaders.

#appsec #aisecurity #devsecops #shadowai #vulnerabilitymanagement #ciso

Rock Lambros, Co-Lead of the 2026 OWASP Top 10 for LLM Applications, on why prompt injection would have fallen out of the top 10 based on incident data alone, and why the community kept it at number one anyway.

In this episode I sit down with Rock Lambros, longtime security leader, former CISO, core team member of the OWASP GenAI Security Project, and Co-Lead of the 2026 OWASP Top 10 for LLM Applications, to dig into the new list and the harder questions underneath it. We get into the letter from the project leads that opens the document, why the LLM and Agentic lists are converging, what a year of messy incident data actually told the working group, and why Rock argues that agency is not authorization.

In this episode:
Why the LLM Top 10 and the Agentic Top 10 are merging in practice
The opening letter, and why the industry was slow to admit the model was never the thing to secure
Who benefits from a model-centric framing of AI security
Prompt injection ranked number one by practitioners and out of the top 10 by the incident data
Why incidents get reported by outcome rather than by initial vector
Misinformation, the widest gap between the vote and the data
Hidden Context Exposure, soft guardrails, and why instructions and data share one context window
Context rot and why context window management matters for agents doing critical work
Using AI to govern AI, and the dual-layer approach in the Agentic Control Standard
Agency versus authorization, and why OAuth is what we have rather than the answer
How to get involved in the OWASP GenAI Security Project
Chapters:
 0:00 Intro and Rock's background
 1:17 What the agentic work changed about the LLM Top 10
 3:59 The opening letter and "It Was Never the Model"
 8:48 Incident data and the prompt injection ranking
 11:23 Misinformation and the limits of the data
 14:24 Hidden Context Exposure and soft guardrails
 17:18 The context window and context rot
 19:57 Using AI to govern AI
 22:49 Excessive agency, and agency versus authorization
 27:09 How to get involved with OWASP

Connect with Rock:
LinkedIn: https://www.linkedin.com/in/rocklambros
OWASP GenAI Security Project: https://genai.owasp.org
Get involved: https://genai.owasp.org/contributing
OWASP Top 10 for LLM Applications 2026: https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/
Read my piece on the new list, It Was Never the Model: https://www.resilientcyber.io/p/it-was-never-the-model
Resilient Cyber: https://www.resilientcyber.io

Ondrej Vlcek, CEO of AISLE and former CEO of Avast, on why AI vulnerability discovery is not as commoditized as the industry thinks, and why remediation is still the real bottleneck.

In this episode I sit down with Ondrej Vlcek, Founder and CEO at AISLE. Ondrej spent roughly 30 years in cybersecurity, joining Avast as employee number six or seven doing kernel-mode driver work on Windows 95, eventually becoming CTO and then CEO, taking the company public and selling it to NortonLifeLock in a nearly $9 billion transaction. He co-founded AISLE in the fall of 2024 to close the loop from discovery through triage, remediation, and verification. His team has now disclosed 350 plus CVEs across projects like OpenSSL and curl.

We get into why the moat sits in the system and not the model, why the gray market price of vulnerabilities has not collapsed even as models get cheaper, and what it actually takes to ship a patch a maintainer will accept.

In this episode:
- Going from Avast intern to CEO, and why vulnerability management was the next problem
- The jagged frontier, and why bigger models do not always mean better results
- Which classes of bugs got cheap to find and which are still genuinely hard
- Why vulnerability prices have not collapsed despite all the model progress
- Building a model-agnostic system with bespoke benchmarks for model selection
- Sovereign AI, on-prem and air-gapped deployment, and why findings are the real crown jewels
- Triage, reachability, and why most findings are not actually exploitable
- Patch verification, regression risk, and mitigations for embedded systems that cannot be patched
- How AISLE earned trust from curl after Daniel Stenberg killed the bug bounty
- Whether a CVE count is a vanity metric
- Build versus buy as model capability keeps getting cheaper
- What breaks first in the CVE and open source maintainer ecosystem
- What AppSec leaders should change next quarter

Chapters
0:00 Intro
0:24 From Avast employee number six to a $9 billion exit
3:26 Why vulnerability management, and why now
5:15 The jagged frontier and what bigger models miss
10:36 The economics of finding bugs, and why prices have not collapsed
12:11 Building a model-agnostic system with real benchmarks
14:30 Sovereign AI, air-gapped deployment, and who sees your findings
19:42 Triage, reachability, and why remediation is the bottleneck
24:48 Patches that break things, and systems you cannot redeploy
25:39 curl, Daniel Stenberg, and death by a thousand slops
29:35 Is a CVE count a vanity metric?
31:34 Build versus buy when capability keeps getting cheaper
34:41 What breaks first in the next 18 months
39:46 What AppSec leaders should do next quarter
41:13 Closing

Ondrej Vlcek on LinkedIn
AISLE
AISLE research and blog

Resilient Cyber Substack
Subscribe for more conversations with security practitioners and leaders.

Lovable CISO Igor Andriushchenko on soft guardrails vs. hard boundaries, securing vibe coding for non-developers, and building a security program at a 10x company.
I sit down with Igor Andriushchenko, Head of Security and CISO at Lovable, the AI development platform behind one of the fastest growth stories in the space. Igor joined as the first security hire when the company was around 40 people. A year later he is running a 20+ person team covering product security, GRC, IT, and platform safety for a company with 400 laptops in MDM and no sign of slowing down.
We get into what it actually takes to secure AI-native development, both inside a hypergrowth startup and on a platform where most of the people shipping software are not developers and definitely not security practitioners.
In this episode:
Building a security program for the company you will be in 12 months instead of the one you are in today
Soft guardrails versus hard guardrails, and how to decide which one a problem deserves
Why hard blocks push AI-assisted workflows into the shadows
Rooting guardrail decisions in business goals, risks, and threats rather than tool defaults
Democratized development without democratized security, and what a platform owes the 99%
Lovable's auto-fix toggle, per-app threat models, and the goal of an app with no security tab at all
Whether models will ever produce secure code by default, and why defense in depth still carries the load
Governing the reality that every employee vibe coding an app looks a lot like a new vendor
GRC engineering as the way to measure control efficiency layer by layer against AI-powered attackers
CRA, NIS2, and the EU AI Act landing on citizen developers who never thought of themselves as software manufacturers
Chapters: 0:00 Intro 0:23 Igor's background from DevOps to CISO 3:54 Scaling security at a 10x company 6:07 Reinventing the team when growth breaks it 08:26 Soft guardrails versus hard blocks 14:05 Tying guardrails to business risk 17:32 Democratized development, undemocratized security 18:52 Shared responsibility on an AI dev platform 21:16 Auto-fix, per-app threat models, and no security tab 25:21 Will models produce secure code by default? 29:56 Every employee vibe coding is a new vendor 30:57 Enterprise controls, publishing gates, and PII scanning 36:39 AI-powered attackers and why good enough changed 40:43 GRC engineering and measuring control efficiency 46:19 CRA, NIS2, and the citizen developer 52:41 Trust centers for builder apps 54:08 Closing thoughts on the vibe coding community
Guest links: Igor on LinkedIn: https://www.linkedin.com/in/igor-andriushchenko Lovable: https://lovable.dev
Resilient Cyber: Newsletter and episode archive: https://www.resilientcyber.io Subscribe for more conversations with security practitioners and leaders.