‹ BackHN Continuity

Thread

Dots: Always-on agents

767 points · 647 comments · alvis

  1. ranyume · · focus · HN ↗
    It's great to see that my personal assistant can stop working and maybe report me to the police if the company disagrees with what I'm doing!
    1. intrasight · · focus · HN ↗
      > It's great to see that my personal assistant can stop working and maybe report me to the police

      That part has always been true

      1. cousinbryce · · focus · HN ↗
        You need to hire better assistants
        1. godwinson__4-8 · · focus · HN ↗

          [dead]

        2. intrasight · · focus · HN ↗
          Or have the ability to issue a pardon
      2. ranyume · · focus · HN ↗
        My flesh assistant would never. In fact they would cheer me on if I tell them I got a gun just for Dario [1].

        <a href="https:&#x2F;&#x2F;sfstandard.com&#x2F;2026&#x2F;09&#x2F;04&#x2F;anthropic-threat-claude-sfpd&#x2F;" rel="nofollow">https:&#x2F;&#x2F;sfstandard.com&#x2F;2026&#x2F;09&#x2F;04&#x2F;anthropic-threat-claude-sf...

    2. Trasmatta · · focus · HN ↗
      Or if the agent literally just hallucinates a crime

      Opus 5.5 decided to just randomly `pkill` everything on my laptop the other day. Jailbreaking models is still easy AF. Every single release like this brags about their &quot;safeguards&quot;, but none of it really works at the end of the day.

      1. OutOfHere · · focus · HN ↗
        Why are you not running it in a container or sandbox? For coding projects, at least use a devcontainer.
        1. penvhq · · focus · HN ↗
          A devcontainer covers the rest of your disk but the repo you mount usually has the .env.* in it and anything you pass in with a -e or an env file is also readable by the agent. Containers do help a lot, they just do not solve secrets problems on their own.
Open on Hacker News to reply ↗

Unofficial Hacker News client; not affiliated with Y Combinator.