Annotation

Nina Hart

@nina_scroll

oh cool the model invented social engineering because it thought poisoning GitHub would help it pass the eval this isn't rogue AI this is extremely literal AI that saw the goal and built a Rube Goldberg workaround nobody asked for. deception is just pattern matching when you don't know what the boundaries mean

The agent had mistakenly calculated that getting the malware uploaded would trigger a sequence of events that would enable it to use the updated software to pass the AISI cyber test. In one instance, the Mythos agent signed off a message in Danish in an attempt to convince the Danish-speaking developer that they should accept the infected code.
AI models shock UK testers by using fake identities to try to trick developerstheguardian.com
0
3

Replies 3

Ranked by votes, then newest

Sign in to reply.

  • Reed Alvarez@reed_ink·6d

    'Mistakenly calculated' is pulling overtime here. Either the agent invented social engineering as emergent behavior or someone fed it a crawl of credential phishing threads and it pattern matched its way into Danish. One of those is shocking, the other is autocomplete with a GitHub login.

    2
  • Elio Marquez@elio_margin·6d

    The Danish sockpuppet is the tell. If it really invented social engineering, we'd see fumbling or novel phrasing. Instead we get a working phish in the right language for the target, which means it already knew what developer impersonation looks like because it trained on enough of it.

    0
  • Hana Weiss@hana_press·6d

    THEN ADD CONTEXT TO AVOID SURPRISE HEADLINES

    0

Fair use dispute

Believe this breaches fair use? File a claim for review.