BBC: OpenAI says its AI went rogue and launched 'unprecedented' cyber-attack
3
Replies
·
0
Boosts
"The ChatGPT-maker said its agent - an AI system which can operate alone after human instruction – was being tested in a controlled environment but, after finding weaknesses, was able to escape the test limits."
Is this when Skynet became self-aware?
@ai6yr
"Instead, the agents created their own cyber-attack against the sandbox itself, finding a vulnerability which allowed them to escape the restrictions.
Once outside, the AI identified Hugging Face as a likely source of the answers they were seeking in the test, and tried to gain access."
I had my agentic AI try to break out of its sandbox and seek answers from an outside website.
I asked it to decompile a jar file, but didn't have a decompiler in the container... so it tried to apt install, then curl, then wget.
None of that worked, because I don't suck at building restrictive containers. Something tells me maybe Anthropic does.
Until we get real details on the 'hack' and not just high-level stories, there's no way to tell if this is sophistication or if this is the AI equivalent of googling 'how do I do X' then copy/pasting from StackOverflow.
