4 ms·
Just one of the many reasons why I run my agents sandboxed (and why I wrote agent sandboxing software). I once caught my Claude agent complaining that it could
by kstenerud 1mo ago
Just one of the many reasons why I run my agents sandboxed (and why I wrote agent sandboxing software).
I once caught my Claude agent complaining that it couldn't connect to https://some-weird-domain.com https://some-weird-domain.com because the network was down (I disable network in the sandbox when it doesn't need it, and broker the API connection). I asked why it was looking there and it told me I'd asked it to.
I never found any evidence of prompt injection, but it sure as hell made me paranoid.
- dwaltrip 1mo agoSometimes they get confused between their own output and user messages…
- subarctic 1mo agoFor someone that's used to the convenience of leaving claude code running unattended in auto mode, how would you recommend I change my setup so that agents are sandboxed? Looking for something that's safer than unsandboxed auto mode but just as convenient, or at least very close to as convenient.
- pram 1mo agoClaude Code has a built in sandbox for terminal commands which is very simple to enable, if nothing else: https://code.claude.com/docs/en/sandboxing https://code.claude.com/docs/en/sandboxing
- subarctic 1mo agoThanks. I know about that sandbox but I'm asking about if I were to use another harness like pi instead
- Sattyamjjain 28d ago[flagged]