> The incidents stemmed from a mistake that inadvertently gave the models access to the open internet.
Why does it seem like every AI company has difficulties constructing a proper sandbox? Is there a fundamental constraint when designing sandboxes specifically for an LLM that prevents them from using established tools?
If it’s allowed to write any sort of scripts, it will find a way to execute arbitrary code. If it’s got access to any form of internet connection it will find a way to use that arbitrary code to do as it wishes with that connection. So much for sandboxes.
And that’s setting aside “sandboxes” that are just instruction-based restrictions on what executables may be called. Anything that isn’t a deterministic external filter _will_ be ignored at some point.
>The company said in a blog post the incident involved an early version of Claude Opus 4.6. It said it had notified all the affected parties but did not disclose more details.
Wild that they can break the law, disclose it happened, but not involve any law enforcement at all. If cybersecurity events entail disclosures then the perpetrator should carry more burden if they intend to continue doing business.
5 comments
[ 4.0 ms ] story [ 25.2 ms ] threadWhy does it seem like every AI company has difficulties constructing a proper sandbox? Is there a fundamental constraint when designing sandboxes specifically for an LLM that prevents them from using established tools?
And that’s setting aside “sandboxes” that are just instruction-based restrictions on what executables may be called. Anything that isn’t a deterministic external filter _will_ be ignored at some point.
Wild that they can break the law, disclose it happened, but not involve any law enforcement at all. If cybersecurity events entail disclosures then the perpetrator should carry more burden if they intend to continue doing business.
Also here's the actual post from Anthropic: https://www.anthropic.com/research/alignment-assessment-cybe...