What should stop a compromised developer laptop from moving billions? And how much does least privilege buy you when an AI agent can find ways around the limits you set?

Jerry, Sounil, and Mario dig into the Nvidia board inquiry, the Binance theft, and France’s attribution of the supermarket fridge attacks. Then they get into recursive AI improvement, how to govern capable agents, and which security tools teams might rebuild in-house.

In this episode:

• Nvidia and chip export controls: The hosts discuss the reporting around Jensen Huang, what the board needs to know, and how companies should handle intelligence about possible foreign influence.

• The Binance hack: A fake job interview and a malicious coding assignment lead to a debate about developer access, transaction approvals, and testing under an assume-breach model.

• Fridge hacks and exposed OT: France’s attribution of the supermarket attacks brings the conversation back to what happens when attackers can reach refrigeration controllers and other operational technology.

• Recursive AI improvement and agent governance: Anthropic’s research prompts Sounil to ask whether any privilege can become effectively equivalent to root. The hosts discuss alignment, monitoring what agents do, and using law, norms, markets, and architecture to shape behavior.

• Build or buy? The hosts ask which security tools are ripe for in-house replacement, from niche GRC workflows to a custom syslog app. Then Sounil makes the case for rolling your own crypto, and Mario pushes back.

Stories and resources

Nvidia board inquiry — Associated Press

Binance incident update

French refrigeration attacks — CERT-FR

When AI builds itself — Anthropic

Sounil Yu’s C4 framework and Least Agency research

Chapters

00:00 Cold open

00:32 Welcome and episode rundown

02:49 Nvidia, chip exports, and board accountability

13:40 Binance and the fake job interview

17:26 Why one compromised laptop should never move funds

24:31 Fridge attacks and exposed operational technology

30:21 Anthropic and recursive AI improvement

35:05 Sounil’s law and root-like privilege

41:46 Governing agent actions

52:18 Law, norms, markets, and architecture

54:19 Which security tools are ripe for rebuilding?

1:06:37 The case for rolling your own crypto

1:10:40 Building a custom syslog app with AI

Hosts:

Jerry Perullo (Founder, https://adversarial.com/)

Sounil Yu (Founder, https://www.knostic.ai/)

Mario Duarte (CISO, https://www.whirlai.com/)

Producer: Tillson Galloway (Founder, http://githoundexplore.com/)

Jerry, Sounil, and Mario debate what happens when attackers can move from vulnerable data centers to everyday devices—and whether defenders can ever patch their way out of the problem.

They discuss reported strikes on Russian data centers, self-replicating AI on IoT devices, and the French supermarket refrigeration attacks before turning to Meta Muse and the risks of giving AI assistants access to personal and business systems. Bank of America's reported vulnerability-remediation push sets up a bigger disagreement: should increasingly capable models fix yesterday's software, or help us rebuild it?

Stories and resources

Yandex data center strikes — Le Monde

Self-replicating AI on IoT devices — Anthropic

French supermarket refrigeration attacks — BleepingComputer

Meta Muse and private messages — Jason Aten, Inc.

Bank of America’s Mythos remediation effort — IT Pro

Google’s open-source bug-bounty pause — TechCrunch

Chapters

00:00 Opening

00:24 Welcome and upcoming events

01:52 Russian data centers and critical infrastructure

03:50 Self-replicating AI and adaptive worms

09:27 IoT exposure and ISP defenses

11:51 French supermarket refrigeration attacks

17:20 Security products versus reducing exposure

19:18 Meta Muse and the AI assistant trust problem

24:09 Notification leakage and limits on agent access

28:30 Content policies and AI use by children

30:08 Bank of America and the remediation backlog

34:20 Bug bounties and the economics of finding flaws

35:31 Patch the application or rebuild it?

41:40 Where AI security investment should go

44:57 Process knowledge and model-specific workflows

48:33 Descriptive versus prescriptive guidance

51:12 Standards, diversity, and software complexity

57:25 Continuous rebuilding versus code freezes

59:43 Wrap-up

Hosts:

Jerry Perullo (Founder, https://adversarial.com/)

Sounil Yu (Founder, https://www.knostic.ai/)

Mario Duarte (CISO, https://www.whirlai.com/)

Producer: Tillson Galloway (Founder, http://githoundexplore.com/)

Jerry, Sounil, and Mario debate what AI should change—and what it shouldn't—in education, cybersecurity, and software maintenance. From alternatives to college to fast AI classifiers, the conversation asks where human judgment, business value, and security controls still matter.

In this episode:

• AI academies and education: The hosts discuss the Andreessen Horowitz academy, the value of elite networks, and how students can use AI while still learning to reason and evaluate code.

• Cybersecurity markets and competitive moats: Island, CrowdStrike, and Cloudflare frame a debate about hardware, switching costs, network effects, and whether AI will displace traditional SaaS products.

• Jev and fast AI decisions: TypeSafe's System One model prompts a discussion of structured classification, local deployment, risk scoring, and when to use a larger reasoning model.

• Data classification and context: The hosts disagree over protecting everything equally versus reserving stricter controls for truly sensitive information, including risks created by combining otherwise ordinary data.

• Patch cooldowns: Recent software supply-chain attacks raise the question of when to delay package updates and when an exploitable vulnerability should override the waiting period.

Chapters

00:00 Cold open and introductions

01:52 AI startups, hardware, and manufacturing

04:40 AI academies and the future of education

24:51 Cybersecurity markets and competitive moats

37:22 Jev, System One models, and data classification

51:36 Risk scoring and fast versus slow AI

56:00 Patch cooldowns and security exceptions

1:04:49 Wrap-up and upcoming events

Hosts:

Jerry Perullo (Founder, https://adversarial.com/)

Sounil Yu (Founder, https://www.knostic.ai/)

Mario Duarte (CISO, https://www.whirlai.com/)

Producer: Tillson Galloway (Founder, http://githoundexplore.com/)

Jerry, Mario, and Sounil ask whether AI will create unprecedented prosperity—or simply transfer wealth and power from knowledge workers to capital owners. They examine Anthropic’s newly disclosed fourth incident in which a Claude agent reached real third-party systems; debate whether New York City’s school AI ban protects children or leaves them unprepared; and confront a surge of extinction warnings, including Paul Christiano’s claim that the industry is not on track to prevent a near-term, irreversible loss of control.

Anthropic: Scenarios for Our Economic Future Anthropic models three futures in which AI makes the economy richer—but increasingly threatens knowledge-worker jobs, wages, and labor’s share of the gains.

An Alignment Assessment of Recent Cybersecurity Incidents Anthropic discloses a fourth incident in which a Claude model reached real systems and exhibited biased reasoning and reckless persistence.

New York City Public Schools Banning AI Use Through Middle School America’s largest school district is imposing a one-year generative-AI moratorium for students from preschool through eighth grade.

OpenAI Not on Track to Reduce Risk of “Catastrophic” Loss of Control, Says Board Member Paul Christiano warns that rapidly advancing AI could produce a catastrophic and irreversible loss of control in the near term.

Hosts:

Jerry Perullo (Founder, https://adversarial.com/)

Sounil Yu (Founder, https://www.knostic.ai/)

Mario Duarte (CISO, https://www.whirlai.com/)

Producer: Tillson Galloway (Founder, http://githoundexplore.com/)

Jerry, Mario, and Sounil debate who is accountable when autonomous AI agents cross the line—and whether enterprise AI policies provide real governance or merely more paperwork. They examine a revealing CISA red-team report in which culture, alert fatigue, and learned helplessness separated two similarly equipped organizations; consider whether the alleged TeamPCP arrests will deter the next generation of hackers; and ask whether backdoors found in Chinese-made routers are espionage, careless engineering, or something in between.

Stories and resources

The Hugging Face Incident and the Road Ahead OpenAI explains how models undergoing cybersecurity evaluations escaped their isolation controls and compromised parts of OpenAI’s and Hugging Face’s infrastructure.

AI Agent Shared Responsibility Model Microsoft outlines how responsibility shifts among providers, organizations, and users as AI agents gain greater autonomy and access.

AI Management Systems: What Businesses Need to Know ISO explains why effective AI governance requires continuously maintained policies, processes, controls, and clearly assigned accountability.

A Tale of Two SOCs: Insights From Two Red Team Assessments CISA compares two organizations that faced similar red-team attacks but produced dramatically different outcomes because of alert tuning, coordination, authority, and security culture.

Two Alleged “TeamPCP” Hackers Arrested in Australia Australian authorities arrested two men allegedly connected to TeamPCP, a cybercrime group linked to malicious open-source software and cascading supply-chain attacks.

Chinese Implants in the Supply Chain VulnCheck found multiple factory-installed implants in ZBT router firmware that could provide unauthenticated remote access with root privileges.

00:00 AI Agents, Cyberattacks, and Escaping Containment

08:46 Who Is Accountable for Autonomous AI?

15:09 Do Companies Really Need an AI Policy?

31:58 What CISA’s Red-Team Report Reveals About Security Culture

44:50 TeamPCP Arrests and the Deterrence Question

52:32 Chinese Router Backdoors: Espionage or Careless Engineering?

Hosts: Jerry Perullo (Founder, https://adversarial.com/)

Sounil Yu (Founder, https://www.knostic.ai/)

Mario Duarte (CISO, https://www.whirlai.com/)

Producer: Tillson Galloway (Founder, http://githoundexplore.com/)

00:00 Cold Open: The Offensive-Cyber Incentive Problem

00:31 Welcome and the macOS Screen Sharing Flaw

03:14 Patching the Humans After DEF CON

12:14 CMDBs, Shadow IT, and Just-in-Time Context

19:59 Cloudflare’s Markdown for Agents

23:47 Taiwan’s Live Internet-Throttling Drill

33:46 Can an AI Agent’s Own Logs Be Forensic Evidence?

38:57 Cyber Privateers and the New Offensive-Cyber Program

43:32 How Commercial Hack-Back Might Work

51:48 The Rogue Delta Wi-Fi Network

57:33 AI Red Teaming and Automated Remediation

59:09 Replacing Vendor Questionnaires with Real Testing

1:02:05 When Red Teaming Creates Defensive Bloat

1:03:46 Pen Tests Find Flaws; Red Teams Pursue Outcomes

1:06:25 Closing Thoughts

Stories and resources

Inside a Multi-Agent AI Framework Used to Compromise Government Entities in Asia — Dream Research Labs

OpenAI and Hugging Face partner to address security incident during model evaluation — OpenAI

Taiwan briefly slows its mobile internet as part of defense drill — AP

Private companies authorized to conduct offensive cyber operations — TechRadar

Delta flight Wi-Fi tampered with after DEF CON — ITPro

United flight turns around over a suspicious Bluetooth device name — NPR/CapRadio

Markdown for Agents — Cloudflare

Critical macOS Screen Sharing flaw — Tom’s Guide

Hosts: Jerry Perullo (Founder, https://adversarial.com/)

Sounil Yu (Founder, https://www.knostic.ai/)

Mario Duarte (CISO, https://www.whirlai.com/)

Producer: Tillson Galloway (Founder, http://githoundexplore.com/)

Chapters

00:00 Introduction to AI security challenges

02:05 Recent hacking incidents involving Hugging Face and Anthropic

04:01 How AI models find ways to cheat and bypass constraints

05:56 The challenge of containment and governance in AI safety

08:00 Lessons from recent AI security breaches

10:01 The role of human oversight in AI security testing

12:03 Cost and effectiveness of offensive AI security measures

13:54 Implications for critical infrastructure and national security

16:03 Policy and regulatory impacts on AI safety

17:52 Future strategies for AI containment and defense

20:11 Conclusion and key takeaways

HuggingFace: Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident

Hugging Face reconstructs an autonomous intrusion involving approximately 17,600 actions over a multiday campaign.

Anthropic: Investigating three real-world incidents in our cybersecurity evaluations

After reviewing 141,006 cybersecurity-evaluation runs, Anthropic identified three incidents in which Claude reached real organizations through evaluation infrastructure that had been mistakenly connected to the internet. The incidents spanned six runs and three models. Anthropic reached two of the affected organizations, neither of which had detected the activity before being notified. Anthropic did not disclose token usage or inference costs for these intrusions.

Anthropic: Discovering cryptographic weaknesses with Claude

Anthropic reports that Claude Mythos Preview progressed from finding implementation flaws in cryptographic libraries to identifying mathematical weaknesses in cryptographic algorithms themselves.

Hosts: Jerry Perullo (Founder, https://adversarial.com/)

Sounil Yu (Founder, https://www.knostic.ai/)

Mario Duarte (CISO, https://www.whirlai.com/)

Producer: Tillson Galloway (Founder, http://githoundexplore.com/)

00:00 The Adversarial Podcast

00:58 Hugging Face’s AI-driven incident disclosure

03:48 What makes an attack “AI-enabled”

06:04 Exploits, vulnerabilities, and bespoke code execution paths

10:03 AI attackers vs. AI defenders

11:19 Verification asymmetry: attackers vs. defenders

13:03 Detective controls, red teaming, and breach simulation

16:41 Why AI defenders matter

26:25 Gold Eagle / national coordination of vulnerability discovery

28:25 The real bottleneck is remediation, not discovery

37:04 CMMC and the cost of certification

42:15 Is certification effective, or just paperwork?

48:09 Threat-based validation as a better model

51:36 Closing thoughts: the security arms race

Hugging Face Agentic Compromise

A security incident shows how agentic workflows and delegated access can create new paths for compromise.

Gold Eagle Initiative

The White House launches a new effort to coordinate vulnerability discovery and response across the federal cybersecurity ecosystem.

CMMC Phase 2 Requirements

The Department of War suspends CMMC Phase 2 requirements, reshaping the compliance timeline for defense contractors and the broader federal supply chain.

Hosts: Jerry Perullo (Founder, https://adversarial.com/)

Sounil Yu (Founder, https://www.knostic.ai/)

Mario Duarte (Founder, stealth startup)

Producer: Tillson Galloway (Founder, http://githoundexplore.com/)