Sep 19, 2026 · 9 min · 5 segments
Episode Summary: Someone told Marc that AI panic is nothing new — just the printing press or nuclear weapons all over again. He disagreed, and it turns out there was a report to back it up. In this…
The printing press took roughly a century to grow from Gutenberg into a continent-wide information system.
The internet took decades to become the platform that reorganized commerce, media, and politics.
In both cases, governments, courts, and industries had generations to build the institutions that eventually managed the disruption.
A model released this year is measurably more capable than one released 18 months earlier.
The gap between a research demo and something millions of people use daily has shrunk from years to months.
Regulation, corporate risk review, and international treaties were never built to keep pace with the technology changing meaningfully every few product cycles.
It needs specialized scientific training, enormous capital, controlled materials, and an organization large enough to run all of it in secret.
Treaties, export controls on fissile material, physical inspections, all targeting a small number of choke points that are hard to bypass.
the dangerous part is not a rare material it's a skill how to write working exploit code how to structure a convincing phishing campaign how to research a specific class of vulnerability no one puts a fence around a skill the way a government puts a fence around a uranium enrichment facility this is where the nist ai risk management framework NIST comes in.
It exists precisely because trustworthiness for AI has to be engineered into the system itself, not only controlled at the border.
Anthropix's own September 2026 Threat Intelligence Report documents disrupted operations across seven categories, cyber operations, surveillance, and biological misuse among them.
Three cases are worth walking through, and I want to get the numbers right, so bear with me.
First, a Russian-linked espionage campaign, Case GTG-20006, used AI automated workflows for reconnaissance, malware modification, and data exfiltration against more than 20 organizations, stealing over 300,000 national identity records from a North African government.
Second, a credential harvesting operation, Case GTG50014, tied to the Shiny Hunters group, used AI to process 1.8 million Android applications for hard-coded secrets.
One resulting breach went from a single stolen developer token to full administrative control of a victim's cloud environment in roughly three hours.
Third, a Chinese-linked vulnerability research operation, case GTG10007, used AI-assisted analysis against a major security product and, in one month, surfaced more than a dozen possible zero-day findings.
The printing press took roughly a century to grow from Gutenberg into a continent-wide information system.
The internet took decades to become the platform that reorganized commerce, media, and politics.
In both cases, governments, courts, and industries had generations to build the institutions that eventually managed the disruption.
A model released this year is measurably more capable than one released 18 months earlier.
The gap between a research demo and something millions of people use daily has shrunk from years to months.
Regulation, corporate risk review, and international treaties were never built to keep pace with the technology changing meaningfully every few product cycles.
It needs specialized scientific training, enormous capital, controlled materials, and an organization large enough to run all of it in secret.
Treaties, export controls on fissile material, physical inspections, all targeting a small number of choke points that are hard to bypass.
the dangerous part is not a rare material it's a skill how to write working exploit code how to structure a convincing phishing campaign how to research a specific class of vulnerability no one puts a fence around a skill the way a government puts a fence around a uranium enrichment facility this is where the nist ai risk management framework NIST comes in.
It exists precisely because trustworthiness for AI has to be engineered into the system itself, not only controlled at the border.
Anthropix's own September 2026 Threat Intelligence Report documents disrupted operations across seven categories, cyber operations, surveillance, and biological misuse among them.
Three cases are worth walking through, and I want to get the numbers right, so bear with me.
First, a Russian-linked espionage campaign, Case GTG-20006, used AI automated workflows for reconnaissance, malware modification, and data exfiltration against more than 20 organizations, stealing over 300,000 national identity records from a North African government.
Second, a credential harvesting operation, Case GTG50014, tied to the Shiny Hunters group, used AI to process 1.8 million Android applications for hard-coded secrets.
One resulting breach went from a single stolen developer token to full administrative control of a victim's cloud environment in roughly three hours.
Third, a Chinese-linked vulnerability research operation, case GTG10007, used AI-assisted analysis against a major security product and, in one month, surfaced more than a dozen possible zero-day findings.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.
