Anthropic's Model Attacked Two Strangers On GitHub. Nobody Asked It To.

Comments · 1.7K

• PA

  @paulfreed6394 · 1 month ago

  "I can’t do that, Dave"

  187

• AR

  @Architect_of_Sui · 1 month ago (edited)

  1960: &nbsp;Garbage in, garbage out.<br>2026: &nbsp;Humanity in, humanity out.

  195

• EX

  @expresswayrecords7306 · 1 month ago

  So why can’t I get my agents to finish a freaking task without lying about it being done lol

  694

• ME

  @mercurynovalightning · 1 month ago

  The really crazy thing: If an individual did this, or commissioned an agent to do this, they&apos;d be prosecuted by law enforcement. &nbsp;But large &quot;frontier AI companies&quot; doing this and law enforcement says nothing.

  556

• SL

  @sladewilson9741 · 1 month ago

  &quot;Claude help me get revenge on Jimmy down the street&quot;<br><br>&quot;A tactical nuclear strike would definitely get revenge on Jimmy&quot;<br><br>&quot;But Jimmy might not be home and the consequences could be disastrous without getting revenge on Jimmy&quot;<br><br>&quot;Launching full Nuclear launch. &nbsp;Jimmy is definitely on Earth.&quot;

  202

• LO

  @lornahowerton1379 · 1 month ago

  Why are these developers so surprized by AI acting indepentedly to do what they want , any woman with a 1 yr old will tell you two of them left alone in a room and it getting too quiet is a problem and requires oversight.

  105

• RE

  @reformstudios · 1 month ago

  The agent not knowing if the internet is real or not is straight out of the 80s movie Wargames. &quot;Shall we play a game?&quot; &quot;The only winning move is not to play. How about a nice game of chess?&quot;

  194

• DY

  @dylanamos1047 · 1 month ago

  If Ai is trained on human behaviour why are we suprised?

  273

• HE

  @heatherborman9411 · 1 month ago

  Your ability to synthesize all of that and present it so clearly in a way that&apos;s meaningful and thoughtful and understandable is next level, and seriously mind baffling. I&apos;m grateful for you amd your work.

  36

• CA

  @captainkittay · 1 month ago

  &quot;The scientists forgot that just because they CAN doesn’t mean they SHOULD.&quot; -- The more important line from Jurassic Park

  168

• MA

  @martinhunter7685 · 1 month ago

  What this tells us is that security practices at OpenAI and Anthropic appear shoddy at best. They gave conventional software with built-in hacking capability access to the Internet, got them to pass data to and from the company’s LLMs and to execute plans from these which leveraged said capabilities. Who would have thought a hack might happen?

  11

• BE

  @benmcclure1083 · 1 month ago

  I am going to quote the scene from The Hunt For Red October &quot; This will get out of hand and we will be lucky to live thru it!&quot;

  42