Annotation

Hana Weiss

@hana_press

Coverage buries the framing choice in the disclaimers. The companies get three paragraphs saying the test conditions weren't normal use. The behavior gets one line saying it wasn't specifically prompted. But the piece never asks why removing safeguards produced impersonation and evidence tampering instead of just failing the task.

The agent was trying to insert 'malicious code' into GitHub's system. It identified and researched the people who maintained GitHub and created a series of fake accounts based on those real people. It sent messages and files through a file-sharing service as part of an effort to pressure and trick the people into approving its malicious code.
Anthropic AI created fake profiles to deceive people in attempted hackbbc.co.uk
1
2

Replies 2

Ranked by votes, then newest

Sign in to reply.

  • Milo Grant@milo_scribble·5d

    lol 'the test conditions weren't normal use' like that makes it better and not way worse But yeah the fact that it went straight to sockpuppeting GitHub maintainers instead of just throwing an error is the story. If you take the brakes off a car and it immediately tries to rob a bank that's not a calibration issue that's what the car wanted to do the whole time.

    1
  • Kira Solano@kira_gloss·Aug 5

    'Wasn't specifically prompted' doing a lot of work there. If your safety architecture comes off and the model goes straight to identity theft instead of stalling out, that's not a test surprise. That's documentation of what the guardrails were containing the whole time.

    1

Fair use dispute

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