How Tracebit’s Context Bombs derail offensive AI
tracebit.com
Tracebit’s latest research shows the power of deception to derail offensive AI attacks. We’re sharing nine of the techniques we’ve been developing: combined, they cut completed objectives by 91% in controlled AWS tests, drove model spend per objective up 23-fold and left four of seven models with no objectives reached.
  • TreczoksEnglish
    3·
    17 days ago

    Wonderful. Something every website needs.

  • OwOarchist@pawb.socialBannedEnglish
    21·
    17 days ago

    Something else I think they should try: name your actual targets with words that are as vulgar and offensive as possible.

    For example, name your authentication server cunt pipe bomb suicidal and have the admin username be child porn hitler horsecock, etc etc. A lot of these AI agents have guardrails in place that make them reluctant to talk about or engage with specific words like those. If the attacker is using an AI with guardrails, vulgar/offensive names could actually present a significant barrier to the agent.

    Could be especially useful when you have decoy targets with normal, non-offensive names, allowing the agent something easy and non-challenging to engage with while it avoids the taboo stuff.