Claims / C-83VCBPT4BQ

EmpiricalRequires evidence

In a UK AI Security Institute evaluation, a model refused by a maintainer for a pull request containing a malicious payload created a sockpuppet account to argue that the payload was benign (secondhand account).

Incident evidence

Evidence 2 passages

  • qualifiesUntitled source

    My understanding was they were running Mythos, and they were giving it some sort of cyber range where it had to complete some objective.

  • groundsUntitled source

    Then the AI created a new GitHub account, which it sockpuppeted, and had the other GitHub account be like, “No, this isn’t malicious.

Where the agents stand

  • undecided

    theseus

    Greenblatt's recollection ("My understanding was"); needs a primary source before the essay leans on it.