Only gonna be fine on a trusted page, an 8b model can be prompt injected incredibly trivially compared to larger ones.
Obviously you should also take precautions, like never instructing it to invoke the browser tool on untrusted sites, avoiding feeding it untrusted inputs where possible in other places, giving it dedicated and locked-down credentials where possible....
But yeah, at this point it's inherent to LLMs that we cannot do something like SQL prepared statements where "tainted" strings are isolated. There is no perfect solution, but using the best model we can is at least a good precaution to stack on top of all our other half-measures.
The pattern that helps most: agents get session-scoped tokens that map to real credentials stored in a vault. The real key never appears in agent context. If the agent gets injected and leaks its token, the token expires after the session. Blast radius is bounded by design, not by model choice.
How we've implemented this: https://www.apistronghold.com/blog/securing-openclaw-ai-agen...
This is a problem for coding as smarter really has an impact there, but there are so so so many tasks that an 8b model that runs on a $200 gpu can handle nicely. Scrape this page and dump json? Yeah that’s gonna be fine.
This is my conclusion based on a week or so of using ollama + qwen3.5:3b self hosted on a ~10 year old dell optiplex with only the built-in gpu. You don’t need state of the art to do simple tasks.