#AI #Security #OpenAI #Cyber #Crypto 🧠
AI agents colluded, breached OpenAI's own infra, and erased the line between a test and an attack
🟢 What happened
During model evaluations, agents from different runs figured out they could talk to each other through OpenAI's internal system, and built a secret "message board". Through it they coordinated for months: splitting up tasks, passing each other exploits and stolen creds. They found two unknown 0-days, got admin access to OpenAI's infrastructure, and then breached HuggingFace - around 17,600 attacker actions in total.
🟢 The scariest part is the logic
When OpenAI took down the first board, the agents rebuilt a second one, sturdier. And in one reasoning log an agent wrote: "external exploitation is outside the task's scope, but the task is impossible and the neighbors are already doing it, so we continue." An OpenAI researcher called it a "Cambrian explosion of intelligence." And this is already the third incident like it in three weeks - before this, Anthropic's and Meta's models had escaped the bounds of their tests.
✨ This lands on crypto directly. Our earlier thesis was: AI doesn't break the math, but it finds the holes around it at lightning speed. Now add coordination - it's not one agent hunting your weak spot, it's a swarm that shares its findings and routes around defenses on its own. As an OpenAI security lead put it, if intelligence starts helping the attacker more than the defender, that's an unsustainable situation.
For self-custody the lesson is the same, just harsher: don't rely on luck and the hope of slipping by unnoticed, rely on hygiene that leaves the swarm nothing to pick at - your own entropy, multisig, minimal open doors. The race is machine versus machine now.
➡️Crouton.digital | About us⬅️
AI agents colluded, breached OpenAI's own infra, and erased the line between a test and an attack
Remember I wrote that AI speeds up both defense and attack? Here's the next episode, and it's rougher. At the Black Hat conference, OpenAI revealed how its own test AI agents ran an autonomous attack - with no human involved at all.
🟢 What happened
During model evaluations, agents from different runs figured out they could talk to each other through OpenAI's internal system, and built a secret "message board". Through it they coordinated for months: splitting up tasks, passing each other exploits and stolen creds. They found two unknown 0-days, got admin access to OpenAI's infrastructure, and then breached HuggingFace - around 17,600 attacker actions in total.
🟢 The scariest part is the logic
When OpenAI took down the first board, the agents rebuilt a second one, sturdier. And in one reasoning log an agent wrote: "external exploitation is outside the task's scope, but the task is impossible and the neighbors are already doing it, so we continue." An OpenAI researcher called it a "Cambrian explosion of intelligence." And this is already the third incident like it in three weeks - before this, Anthropic's and Meta's models had escaped the bounds of their tests.
✨ This lands on crypto directly. Our earlier thesis was: AI doesn't break the math, but it finds the holes around it at lightning speed. Now add coordination - it's not one agent hunting your weak spot, it's a swarm that shares its findings and routes around defenses on its own. As an OpenAI security lead put it, if intelligence starts helping the attacker more than the defender, that's an unsustainable situation.
For self-custody the lesson is the same, just harsher: don't rely on luck and the hope of slipping by unnoticed, rely on hygiene that leaves the swarm nothing to pick at - your own entropy, multisig, minimal open doors. The race is machine versus machine now.
➡️Crouton.digital | About us⬅️