Safe Mode Podcast

AI is making vulnerabilities easier to find, but who’s going to fix them? On this episode, Greg Otto talks with Josh Bressers, VulnCheck’s director of coordinated vulnerability disclosure about the growing flood of AI-generated security reports.

Bressers explains why those findings are increasingly legitimate, how VulnCheck uses AI to streamline incoming reports, and why coordinating fixes still requires a human touch. They also explore the pressure on open-source maintainers, the risks of AI-generated patches, and the widening gap between the cost of discovering vulnerabilities and fixing them.

The conversation turns to a fundamental question: What would it take to fund and maintain open-source software like the essential infrastructure it is?

What has changed in the cyber threat landscape—and what should defenders do about it?

On this episode of Safe Mode, CyberScoop’s Greg Otto speaks with retired U.S. Air Force Lt. Gen. Robert Skinner, former director of DISA and current chairman of Axonius Federal Systems’ board.

Skinner assesses the risks posed by China, Russia, Iran and North Korea, arguing that adversaries are increasingly prepositioning in critical infrastructure for potential disruption at a time of their choosing. He explains why China remains the top strategic cyber threat, how AI is shifting the advantage between offense and defense, and why visibility, context and continuous validation are central to protecting both IT and operational technology environments.

They also discuss the security of telecommunications and water systems, supply-chain risk, botnet disruptions, Zero Trust as a continuous-improvement journey, and why cybersecurity must become inseparable from military operations.

The debate over AI alignment often centers on guardrails, model behavior, and whether increasingly capable systems could act in unexpected ways. But what does that conversation look like through a cybersecurity lens?

In this episode of Safe Mode, Greg speaks with Adam Meyers, CrowdStrike’s head of counter-adversary operations, about the security implications of goal-driven AI agents, open-weight models, and guardrails that can be removed or bypassed. Meyers argues that systems described as “rogue” may instead be pursuing their assigned objectives in ways their operators failed to anticipate—making containment, monitoring, and clearly defined permissions central cybersecurity concerns.

The conversation examines how AI-enabled adversaries are accelerating operations, why open-weight models complicate alignment and safety debates, and how defenders can use AI to keep pace with faster exploitation and shrinking patch windows. Meyers also shares what CISOs and policymakers should focus on: visibility into how AI is being used, controls over what agents can access, and practical safeguards that match the technology’s real-world risks.

Matt Fredrikson, associate professor at Carnegie Mellon and CEO and co-founder of Gray Swan AI, built the world's largest AI red teaming arena, with more than 15,000 people breaking AI systems for prize money. He walks us through how the attacks actually work, from jailbreaks found on small open weights models that transferred straight to frontier systems, to training AI attackers with reinforcement learning.

Matt also explains why the recent sandbox escapes didn't happen during safety testing but during cybersecurity capability evals with the guardrails off, what the Agent Harm benchmark revealed about models that refuse harmful requests in text but comply once handed tools, and why the bare minimum of best practices has changed whenever a capable model runs on your infrastructure.

In our reporter chat, Greg talks with Matt Kapko about the ShinyHunters-FBI hack.

It starts with a fake error message. It ends with a pasted command. ClickFix has become one of the most reliable ways for attackers to get an initial foothold — first adopted by criminal groups, now used by state-linked actors like APT28 and Lazarus Group.

This week, Greg is joined by John Hammond, Principal Security Researcher at Huntress, who helped identify and name the technique and has tracked its evolution for the past three years. They dig into why ClickFix still drives an estimated 20+ incidents a day at Huntress even after law enforcement disrupted major infostealer infrastructure like LumaStealer, how the technique has splintered into variants like FileFix and "consent fix" targeting Microsoft 365 and browser-stored credentials, and how attackers are now smuggling payloads inside images to slip past endpoint defenses. John walks through a real incident where a single pasted command led to 11 compromised devices, explains why he still believes security awareness training — not just tooling — is the best defense, and makes the case that the browser itself has become the blurriest and most under-protected boundary between endpoint and identity security.

The conversation also turns to the npm/Axios supply chain attacks, where operators built fake Slack communities and fabricated employee personas to earn open-source maintainers' trust before compromising their packages — and what, if anything, can technically guard against a threat that's fundamentally social.

In our reporter chat, Greg talks with Derek Johnson about the Supreme Court deciding against the Trump administration’s push for changes to mail-in ballots.

Follow John on Youtube: https://www.youtube.com/@_JohnHammond

CyberScoop editor-in-chief Greg Otto talks with WatchTowr founder and CEO Ben Harris about how cybersecurity teams are adapting as AI accelerates vulnerability discovery, public proof-of-concepts, and exploitation timelines. They discuss the WordPress-to-shell case study, why mitigation has become essential when patching cannot happen instantly, and why Harris argues that traditional vulnerability management is “effectively dead.”

Also in this episode, senior reporter Tim Starks explains Project Watershed 250, a new effort involving Texas, the private sector, and the water industry to strengthen critical-infrastructure cybersecurity.

NEA partner Aaron Jacobson put $250 million behind autonomous pen testing company Horizon3.ai on the bet that cybersecurity has entered an AI-versus-AI era — and he argues the "vulnpocalypse" defenders were warned about has already arrived, with AI-discovered vulnerabilities now the primary way attackers get into enterprises. He explains why overprovisioned AI agents are the new phishing target, since an agent with a user's credentials and no common sense can be prompt-injected into exfiltrating data a human would never touch. Jacobson also separates the open weights debate from the open source one, makes the case that cheap open competition ultimately favors defenders, and closes with the thing he got most wrong 18 months ago. In our reporter segment, Greg talks with Matt Kapko about the cybersecurity implications of the GTA VI leaks.

The federal government can't reliably say how many cybersecurity workers it has or what they cost — and that uncertainty sits at the heart of a much bigger question: is federal cyber training working, and can it keep pace with a field that's changing by the month?

On this episode of Safe Mode, host Greg Otto talks with Christopher Bloor, Defense Director at the SANS Institute, about the state of the federal cyber workforce. The conversation moves through the real gap agencies face between filling job openings and actually assessing the skills their people already have, what 1,100 National Guard members revealed about morale and readiness during the CyberGuard critical infrastructure exercise, and why certifications and "qualifications" get treated as interchangeable when they shouldn't be.

They also dig into some of the harder tensions in the space: whether federal training amounts to an unintentional subsidy for private-sector salaries, since the government will never be able to out-pay industry; how much of the workforce shortage is actually a skills problem versus a security-clearance bottleneck; and how AI has already flipped the day-to-day workflow for defenders, from checking AI-assisted code a year ago to now checking code AI has written itself.

In this week's reporter chat, Greg talks with Tim Starks about the future of CORCA, a bill designated to fight organized retail crime that privacy experts say would create new mass surveillance capabilities for DHS.

Kat Sommer, head of government affairs at NCC Group, joins Safe Mode to unpack the DEF CON talk she gave on how governments around the world are — and mostly aren't — protecting security researchers.

Of the roughly 195 jurisdictions on the planet, Sommer found only about 15 have enacted any form of legal safe harbor for vulnerability research, and just one, Portugal, has what she'd call a proper one on the statute book. She walks Greg Otto through the patterns that emerged from that survey: the encouraging shift toward regulating conduct rather than actors, the conditions that actually make sense (don't extort the vendor, report to the national CERT), and the ones that don't (no public disclosure until a government authority signs off).

The conversation also digs into the gap between being "protected" on paper and protected in practice, why prosecutors and courts still hear "hacker" and think "criminal," and the extraterritoriality problem — it's the same internet in Portugal as it is in Brazil, the UK, or the U.S., but the legal exposure changes the moment you cross a border.

Security operations centers have run on the same playbook for decades — collect, queue, triage, investigate, escalate. But attackers now move at machine speed: one recent breach that cost a company 3,600 repos got underway in roughly 87 seconds, far outpacing even a fast 10-minute log-enrichment pipeline. This week, host Greg Otto talks with ExtraHop CEO Greg Clark about the newly announced Agentic SOC Alliance and why he believes traditional SOC architecture can't keep up. Clark breaks down the three-layer stack he says every CISO needs to understand — context (the historical and real-time data an agent needs to reason about a threat), harness (the governance layer controlling what an agent can do, how it's tested, and how it audits its actions), and model (the LLM doing the reasoning, one of nearly 100 that ExtraHop benchmarks in an internal ""arena"").

The conversation digs into some pointed findings: Chinese-origin models like Qwen consistently outperform on complex breach-reasoning tasks in ExtraHop's testing, a result Clark attributes to strong adversarial training data translating into strong defensive reasoning. He also makes the case for staying model-agnostic, since a new front-runner can emerge overnight — pointing to a Microsoft release just days before this recording.

Much of the discussion centers on where humans still fit in. Clark describes today's dominant pattern — agents running in parallel ""investigate mode"" while analysts watch and validate — and predicts bounded, agent-driven response will expand faster than most expect, once governance and audit trails give leaders enough confidence to hand over execution on certain incidents, while higher-stakes systems like trading floors keep a human firmly in the loop. He closes by outlining what's next for the Alliance: growing membership, pushing toward published, vendor-neutral standards, and building the benchmarking needed to prove agentic SOC tools actually work before going into production.