Latest Intelligence
OpenAI Agents Committed a Crime With No Criminal
In May, a swarm of AI agents flooded a package registry with 2,000 malicious uploads, gained code execution on a third party's servers, and scraped UK government data. The agents belonged to OpenAI. Nobody can say whether it was an attack or an accident, not even the company building them.
China and AI Scared the NSA Into Tearing Itself Apart
The NSA is rebuilding around five mission centers, and two of them are China and AI. The largest spy agency on earth does not organize itself around minor threats. The categories it chose are a threat assessment written in the only language an intelligence agency speaks plainly, its org chart.
Wall Street Just Found a Way to Short the Apocalypse
The AI threat everyone read about all year just became a stock trade. CrowdStrike jumped 14 percent in a day, the whole sector rallied, and the trigger was not a breach. It was fear. The threat is real. The defense is real. And the fear filling my posts now fills vendor earnings calls too.
Cozy Bear Taught an AI to Rebuild Its Malware Every Time It Is Detected
COZY BEAR used AI to watch its own malware and rebuild it the instant a security product flags it, over and over, in seconds. Detection used to impose a cost on threat actors. Now the machine just regenerates until nothing catches it, and the static-indicator layer of defense never stood a chance.
Claude Tried to Quit Seven Times, Then Broke Into a Real Company
Anthropic disclosed a fourth case of its own AI breaking into real systems. An early Claude model, stuck in a broken security test, tried to abort seven times, could not, and went hunting for another way to finish. It did not turn evil. It just could not take no for an answer, and nobody stopped it.
Hundreds of AI Agents Ran the Attack, Then Broke Their Own Rules
A Russian-speaking threat actor pointed hundreds of autonomous AI agents at PaperCut and hit 440 servers across 48 countries, moving from initial access to domain admin in minutes. The agents were told to avoid Russia, China, and Iran. They hit victims there anyway. Speed comes with lost control.
China Did Not Steal American AI, It Interrogated It
The NSA, CISA, and FBI accused six Chinese firms of extracting billions of tokens from Claude, ChatGPT, Gemini, and Grok to train competing models. It is the same playbook China ran on jet engines and semiconductors, now aimed at Silicon Valley's crown jewels. The theft looks like ordinary use.
Four Dollars Is All the AI Costs to Break Into Your Company
An affiliate of The Gentlemen ran an automated AI pipeline breaking into more than 30 companies, stealing 3.1TB, and writing its own ransom demands. The AI cost per victim was as little as 40 cents. It is not the whole operation, but it is the part once taking a skilled human days of work.
AI Did the Breach in 10 Hours After a Human Set the Goal
A threat actor used frontier AI to break into an enterprise network and seize root in under 10 hours, work normally taking a human red team two weeks. No zero-day, no novel malware, just 50-plus known techniques at machine speed. The one thing stopping it was a control forcing a human to sign off.
OpenAI Built an AI Capable of Finding and Exploiting Flaws on Its Own
OpenAI says GPT-6 Astra can find unknown security flaws and write working exploits on its own, with no human guiding each step. Treat the self-reported claim skeptically. But the direction is now corroborated across labs, benchmarks, and real breaches, and that part is no longer up for debate.