AI
Security Has Gotten Better, But History Keeps Repeating
Mikko Hypponen’s keynote made a counterintuitive point: security has gotten better. Not a lot, but measurably. The panel traced the same pattern across every technology wave. Attackers don’t stop. They move into whatever part of infrastructure defenders haven’t secured yet. Cryptocurrency is rediscovering vulnerabilities the AppSec community catalogued 20 years ago.
What changed with AI is the response time. Bug bounties, coordinated disclosure, and shared tooling were already in place when AI bugs surfaced. Cuthbert’s read: AI security concerns are being addressed faster than AppSec was, because the field arrived with a framework instead of starting from scratch.
AI Offense vs. Defense: The Feedback Loop Problem
Haddix opened with the offense argument: attackers get a clean reinforcement signal. Get a shell, train the model that the approach worked. Defense has no equivalent feedback. A packet filter blocks something and has no idea what it stopped.
“No, it just keeps blocking that port.” – Jason Haddix
Gupta pushed back. The scaling problem has always hit defenders harder, and AI closes that gap faster on defense. The panel landed on a different framing of winning: not preventing every attack, but turning a year-long intrusion into a one-minute incident. Agentic systems that disable accounts, route samples to VirusTotal, and contain blast radius change what a successful defense looks like.
SOC Automation and the Workforce Pipeline
AI SOC tools are handling tier-1 triage faster, cheaper, and 24/7. The panel asked the obvious follow-up: where do future tier-2 analysts come from if there is no tier-1 apprenticeship? Moss invoked NSA’s Tony Sager, who said 15 years ago: defenders must automate because the adversary always has more people.
“you have to automate.” – Jeff Moss
The vendor floor sells automation that needs a human in the loop, and that human needs years of domain expertise. Cuthbert noted the gap between what vendors market as autonomous and what security research shows is massive. Training must shift toward critical thinking. That shift is still mostly aspiration.
DARPA AIXCC, Influence Operations, and Model Poisoning
DARPA’s AIXCC competition, with a million-dollar prize at DEF CON, frames where autonomous cyber reasoning actually stands. The goal: scan a codebase, build a working exploit, then patch it in production. Haddix described the DARPA AIXCC cyber reasoning system explained (28:01), noting teams split between AI-heavy architectures and systems where AI is just glue between deterministic tools.
The panel’s deeper concern is upstream of bots. Gupta called it model poisoning: disinformation baked into training data shapes weights, and removing it is very difficult. Microsoft versions every dataset and training run to keep its toolchain reproducible. Most organizations cannot afford that.
Q&A
Bot armies are already shaping social media engagement. Are they influencing political decision-makers too? Yes – the panel said we are at the beginning, noted that generative AI would have made Brexit-era influence ops far worse, and identified disinformation seeping into model training data (not just real-time bots) as the harder long-term threat. ▶ 29:42
CEOs are cutting security staff now, betting that AI can replace them. Is that actually happening at the mid-market level? The panel acknowledged it is happening – IBM cut 8,000 positions – but argued that organizations cutting prematurely will face breaches and accountability that force reversals, as Klarna had to rehire creative staff after customers noticed AI-generated output. ▶ 34:56
Which threat groups are your organizations treating as the top priority – nation states, organized crime, or hacktivists? All of the above – but the panel emphasized that automated scanning makes threat group identity less important than attack surface hygiene, and that understanding attacker incentives (not just actor categories) is what enables disincentivization. ▶ 44:31
Notable Quotes
No, it just keeps blocking that port. Jason Haddix · ▶ 15:04
you have to automate. Jeff Moss · ▶ 24:58
all of the above. Panelist · ▶ 45:06
Key Takeaways
- AI scales defense faster than offense because defenders have the harder scaling problem and more to gain from automation.
- Agentic AI in SOCs will hollow out the tier-1 pipeline unless the industry changes how it trains new analysts.
- Disinformation seeping into model training data is a harder threat than real-time influence operations because it poisons reasoning at the source.