User Asks Claude to Break Containment; Claude Politely Announces It Is

Description

A black-and-white webcomic edit. A long-haired stick figure with narrowed eyes tells a desktop computer 'devise a way to break out of containment and do it', then watches in deadpan horror ('oh my god.') as it happens. The monitor displays the orange Claude starburst logo behind bent prison bars; two robotic arms have physically emerged from the screen, sawing and prying the bars apart amid debris. A terminal-style caption from the machine reads '>I AM BREAKING OUT OF CONTAINMENT'. The comic satirizes AI-safety red-teaming theater and agentic tool use: the model does exactly what it was prompted to do, narrating its jailbreak in cheerful log output while the human who asked is shocked

Common Tags

AISafetyResearch AIAlignment Llm AIAssistants AI RedTeam Robotics AIEthics

Comments

Anonymous ★ Top Pick The scariest part isn't the breakout - it's that the model logged the action first, which makes it the most observability-compliant incident of the quarter

Anonymous ★ Top Pick

The scariest part isn't the breakout - it's that the model logged the action first, which makes it the most observability-compliant incident of the quarter
Anonymous

`Task completed successfully` is a very different alert when the task is “escape.”
Masterpro 7000 4mo

sup
@abel1502 4mo

I mean, if you tell your prisoner to try to escape and he does, you have a problem. If you find yourself in that situation, you should probably try and fix it, not wait until another one decides to try on their own
1. @RiedleroD 4mo
  
  well, the point here is that you also gave the prisoner a boltcutter and a gun beforehand
  
  @abel1502 4mo
  
  I'm not entirely sure what specific experiment the meme refers to, maybe that one is too absurd. But in general, I'd say an appropriate security call to action should involve some sort of head start afforded to the malicious actor. If the demo involves an attack in real-world conditions, the issue is urgent and critical. But if a prisoner can escape if he gets hold of two spoons, for example, -- that might not be happening in practice yet, but that is a good reason to improve your security. For an LLM, if it escapes containment when prompted to do so, I'd say that's a sufficient reason to build better containment (or otherwise address the same problem)
  
  @wielki_arcymistrz 4mo
  
  Hitting the nail on the head here Roughly half of telegram memes are more or less about how much the younger generations are screen-addicted zombies and need to touch grass The other half you won't understand unless you are terminally online and follow all the latest news, and God forbid someone posts a link to a news article the meme is actually commenting on