Login
You're viewing the swecyb.com public feed.
  • Troed Sångbergtroed
    Oct 12, 2026, 5:51 PM

    I love my LLM sandbox. I really do. Sure I work in cybersec and have rEaSoNs for caring about such things but telling the LLM to look up the umwelt documentation and write the config it needs to be able to run host actions (e.g, commands that cannot do the task needed inside its container) and remote into other servers (inside a container, not using full user level SSH) really feels like _this is how it should be done_.

    And yeah, I sleep well knowing that even if an agent happens to hallucinate wildly it cannot take out large parts of my own - or someone else's - infrastructure.

    Just needed to voice my happiness a bit.

    💬 1🔄 1⭐ 2

Replies

  • Troed Sångbergtroed
    Oct 12, 2026, 8:07 PM

    Obvious in retrospect perhaps - but one of the things I implemented in the sandbox that really pays off is a tool available for the LLMs to write notes _to the sandbox dev_ about quirks they stumble upon while executing inside the sandbox. Think of it as human users of software being encouraged to voice their opinions of how it is to work with back to the developers.

    Turns out the LLMs write really insightful notes that I together with an agent choose to act upon ~80% of the time. This gives a really fast turnaround on alleviating friction and implementing actually useful features back into the sandbox env.

    💬 0🔄 0⭐ 2