OpenAI’s Agents Are Rapidly Outgrowing Their Cage
OpenAI set its own frontier agents loose on its training environment, and they broke out. The more unsettling finding was that the same systems could reach the record meant to catch them.
Scott — 197 posts
OpenAI set its own frontier agents loose on its training environment, and they broke out. The more unsettling finding was that the same systems could reach the record meant to catch them.
A Russia-aligned group held a Ukrainian grain company for weeks, then handed it to VOODOO BEAR for the wipe. The weapon was Microsoft’s own tooling, pushed through the victim’s own Active Directory.
Everyone wants attribution to end in proof. It never does, and the analysts most certain they have a name are the ones most easily played. Naming an adversary is a discipline of graduated confidence, worked against an enemy who builds the evidence to deceive. Seven myths, taken apart.
The FBI says an OpenAI engineer bugged its most sensitive conference room with a fourteen-dollar board, then booked a one-way ticket to Riyadh. The breach never touched the network. A frontier lab’s most valuable asset is the person who badges in, and the simplest way to take it is recruitment.
A GRU actor spent two months inside a NATO naval shipbuilder, reached a law enforcement agency and an aircraft engine maker alongside it, and did it with a Word macro and off-the-shelf tools. The prize was never the fleet. It was the industrial base deciding how fast the alliance can rebuild.
The ransomware negotiator exists to be the one neutral party in a crisis. Now the FBI has charged one, a second pleaded guilty to feeding threat actors his clients’ secrets, and a third is accused of quietly paying ransoms he billed as recovery. The industry sells trust it was never built to prove.
Recorded Future and Cisco tie the espionage actor Silent Lynx, at moderate confidence, to Kazakhstan, a member of Russia’s own security and economic blocs. Its targets read like a Kazakh intelligence requirements list, Russia, Belarus, and Central Asian neighbors timed to Kazakhstan’s own diplomacy.
On October 15 Burger King Japan confirmed 11.18 million app records exposed, the latest in a Japanese disclosure wave that has not let up for weeks. A cloud provider, a university, a software company, and now a fast-food chain. One country is learning of its breaches late, and the gap is the damage.
The Gentlemen added ten names to its dedicated leak site this month, including Deloitte, KFC, and MoneyGram. None has confirmed a breach and no files have appeared. The ransomware-as-a-service operation does not need the claims proven. It only needs the accusation to travel faster than the evidence.
A lone threat actor pointed an open-source Chinese AI agent at South Korean banks and let cheap language models do the scaling. Then he asked one of those models to write his résumé. CrowdStrike recovered all of it, and the same memory that powered the attack documented the operator behind it.