Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
this is really bad...
TL;DR
AI is already helping find real zero-days — Matthew Berman points to Google Threat Intelligence Group’s report that a threat actor used AI to develop a previously unknown exploit, with Google saying its own “proactive counterdiscovery” may have prevented a mass attack.
The current wave is supply-chain chaos, not sci-fi AGI doom — he highlights the Shai-Hulud npm worm spreading across 169 package names and 373 malicious package versions, plus crossover into PyPI, as evidence that AI-assisted attacks are happening now and getting more severe.
Vibe coding is widening the attack surface fast — Berman says more code is being shipped, fewer people are reviewing dependencies, and attackers are using AI to generate malware, evade defenses, and scan open-source codebases at machine speed.
Frontier labs are building cyber-defense models almost as fast as attackers are escalating — Anthropic’s Mythos reportedly found a 27-year-old OpenBSD bug and a 16-year-old FFmpeg flaw, while OpenAI’s GPT-5.5 Cyber scored 81.9 on CyberGym versus Mythos at 83.1.
His core thesis is “my AI versus your AI” — borrowing Jensen Huang’s framing, Berman argues the strongest defensive models will likely belong to big labs and states with the most compute, money, and researchers, making elite defense stronger than rogue attackers over time.
The near-term danger is the long tail — even if top companies can harden their systems with models like Mythos or GPT-5.5 Cyber, AI makes it newly profitable to attack smaller teams, indie developers, and ordinary people who have weak defenses and lots of exposed software.
The Breakdown
A clickbait title, then a real warning
Berman opens by admitting the title is “slightly clickbaity,” then immediately says he can back it up: cyberattacks are getting more frequent, more severe, and increasingly AI-supported. He frames this as the first time, after years of covering AI, that he’s actually worried.
Google’s zero-day report changes the tone
The first big proof point is Google Threat Intelligence Group reporting the first known case of a threat actor using AI to develop a zero-day exploit in the wild. Berman explains why that matters in plain English: zero-days are normally hoarded, expensive, and strategically deployed — so AI moving into that territory is a major shift, even if Google’s own AI may have helped stop it.
Shai-Hulud and the npm worm mess
He then moves to the Shai-Hulud supply-chain attack, describing it as “basically taking over the internet right now.” The details are ugly: a malicious npm campaign with a dead man’s switch, GitHub token theft, and a watcher that wipes your home directory if you revoke the stolen token, eventually expanding to 169 package names and 373 malicious versions.
Why AI is making this worse right now
Berman cites a post breaking the trend into two forces: way more code is being written, and attackers have “woken up.” He’s candid that he barely reviews much of his own AI-written code, then connects that casual vibe-coding culture to a much bigger attack surface attackers can exploit with AI at scale.
Vercel, phishing, and the human weak point
The Vercel breach becomes his next example, with CEO Guillermo Rauch saying the attackers moved with “surprising velocity and in-depth understanding of Vercel” and were likely “significantly accelerated by AI.” Berman broadens that to phishing and deepfakes too, repeating advice from PinDrop Security’s CEO: create a family passphrase, because a FaceTime scam using your cloned voice and face is no longer far-fetched.
Mythos, GPT-5.5 Cyber, and the cyber-model race
The middle of the stream is a tour of the defensive side: Anthropic’s unreleased Mythos, OpenAI’s GPT-5.5 Cyber, and OpenAI’s new Daybreak initiative. Berman notes Anthropic gave Mythos to companies like AWS, Apple, Nvidia, and Palo Alto Networks, then calls its “too dangerous to release” posture partly fear-based marketing, especially since OpenAI shipped a comparable model and “the world did not end overnight.”
His big argument: bigger model wins
This is where he leans on Jensen Huang’s “my AI versus your AI” framing and really commits to it. His argument is economic: world-class models need massive compute, energy, data centers, and talent, so random malicious groups won’t outbuild major labs or states — meaning stronger defensive AI should beat weaker offensive AI most of the time.
The long-tail risk is where the pain lands first
Still, he says the interim period is “really bad” because AI makes lower-value attacks newly profitable. His hand-drawn graph moment lands the point: attackers don’t just need giant enterprise jackpots anymore — with AI automation, they can profitably target smaller companies, open-source maintainers, and regular people in parallel, while state-vs-state AI cyber competition remains the part that makes him most nervous.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
