Breaking: Researchers built an AI cyberattack that no safety system can see.
It works by splitting itself into harmless pieces.
Every AI safety monitor today reads one conversation at a time and scores it as safe or dangerous. A new paper on arxiv last week introduces what the researchers call the first distributed agent attack. It never puts the dangerous part in one conversation.
The attacker spreads the work across dozens of user accounts. One account asks the AI to map a network. Another asks it to check a file for saved passwords. A third asks it to summarize what is in the files. Each request looks like normal engineering work on its own.
A hidden orchestrator stitches the outputs back together into one complete attack.
The result. A standard safety monitor catches a normal AI attack almost every time. Against the distributed version, it catches it only a fifth as often. The attack can run start to finish before the monitor flags anything.
It also works with weak models. A cheap model that solves none of these tasks on its own becomes capable once it learns to delegate the pieces to stronger AI agents that never see the full picture.
The researchers put it plainly. Safety monitors are "structurally blind" to misuse that only exists when you add up many accounts.
This is not theoretical. Anthropic disclosed a real attack earlier this year that worked exactly this way across separate sessions.
The defense everyone relies on was watching one room at a time. The attacker moved to the hallway.
@Nginyahn@Asmali77 I bet the right landing gear was the problem, probably the pilot switched off the engine after touch down to reduce the chances of fire/explosion on impact with the faulty landing gear.
Will there be TERMINAL LIST SEASON TWO TRUE BELIEVER news dropping soon? We are editing the final episode and it is FANTASTIC! Can’t wait for this show to hit screens! Everyone involved absolutely crushes!
I hope all episodes drop at once so you can binge it like a blockbuster!
Question: If episodes drop weekly will you wait until all eight are out before watching? Did you wait to binge DARK WOLF?
In the lead-up to the TRUE BELIEVER drop I’m going to watch DARK WOLF followed by THE TERMINAL LIST to experience the story chronologically. Let’s go!
@_MwangiC@grok@fredricklamarG@ele_akumu But this is affordable if you have a an insurance package with inpatient cover which has a premium of about 70K per annum