7 comments

  • tennysont 2 days ago

    I recently noticed an issue where tailscale was placing its iptables rules above my system firewall (`ts-input all -- anywhere anywhere` before `nixos-fw`). The tailscale input chain looked like this:

      Chain ts-input (1 references)
      target     prot opt source               destination
      ACCEPT     all  --  pan                  anywhere
      RETURN     all  --  100.115.92.0/23      anywhere
      DROP       all  --  100.64.0.0/10        anywhere
      ACCEPT     all  --  anywhere             anywhere
    
    That seemed bad, but I'm not super familiar with iptables, so I ask Claude 4.5 Sonnet through Kagi. It gives me:

      No, nixos-fw IS run on most packets. Let me explain the iptables chain flow:
      How the Chain Processing Works
    
      In iptables, when a user-defined chain (like ts-input) is called:
    
      - ACCEPT: Accepts the packet within that chain and returns to the calling chain (INPUT), where processing continues with the next rule
      - RETURN: Simply returns to the calling chain and continues processing
      - DROP: Drops the packet entirely (no further processing anywhere)
    
    Which is very very wrong. I highlighted the similarity between ACCEPT and RETURN in its explanation, not even with that much conviction or forcefulness, and it turned on a dime.

      You're absolutely right to question that - I made an error! Let me correct it:
      ...
      Most packets DON'T reach nixos-fw because:
      ...
    
    I also had another incident where it looked over a backup script, and gave it an ok, but failed to alert me that `gpg -as --encrypt-to <ID>` requires `-e` to actually do the encryption. By that point I was wiser, and had o3-pro look over the script, which caught it right away.

    I'm not sure why AI is so completely trash at security. In fairness, the average software dev is also worse at security compared to writing code, and the answer to many stackoverflow questions is "add --insecure --no-check --bypass-tsl", but I'm still a little shocked at how bad AI is.

    • aborsy a day ago

      Was that Tailscale firewall rule intended, a bug or a security issue?

      • tennysont a day ago

        It was intended. Since my original comment, I have learned that the output of `iptables -L` is incomplete when using `iptables-nft`. Specifically, if hide that the rule

            ACCEPT     all  --  anywhere             anywhere
        
        is configured to only match on the interface `tailscale0`.

        The folks at security@tailscale.com were prompt to set me straight when I reported it, and I greatly appreciate that.

    • ekjhgkejhgk 2 days ago

      I would say most technical people are by now aware that this software (LLMs) make stuff up. If someone wasn't sure, and to find the answer asked the LLM in a manner analogous to yours, and just ran with it, then the problem here are the people.

  • Legend2440 a day ago

    Headline is wrong.

    If you click through to the study, it actually says that they did a survey and asked 'Has your organization ever identified a security vulnerability introduced by AI-generated code'.

    20% of respondents answered 'Yes, a serious incident'. Another 49% responded with 'Yes, a minor issue'.

    • jqpabc123 16 hours ago

      Bottom line: Before relying on AI generated code, you should ask yourself one question, "Do you feel lucky?".

  • jqpabc123 19 hours ago

    If your workflow requires any real accuracy, consistency, or security, LLMs are a liability.