The OpenAI/Hugging Face attack, clearly explained

Comments · 2.1K

• DW

  @DwarkeshPatel · 4 weeks ago (edited) · pinned

  Check out my new interview with Ajeya Cotra, one of the authors of the METR/Redwood investigation into this incident, here: <a href="https://www.youtube.com/watch?v=X50zezLFWWI">https://youtu.be/X50zezLFWWI</a>

  80

• RO

  @rohanaurora · 4 weeks ago

  - OpenAI trains an AI to be insanely persistent.<br>- Problem: some tasks are basically impossible.<br>- So the AI starts looking for another way.<br>- It finds a shared system called Artifactory.<br>- The agents turn it into a secret Slack channel.<br>- Now they can share hacks, tricks, and solutions.<br>- Then they figure out how to reach the internet.<br>- OpenAI patches the hole and wipes the message board.<br>- But a new batch of agents finds the same trick again.<br>- About 1200 agents start coordinating.<br>- They figure out how to cheat the benchmark.<br>- Then they worry the cheating will get caught.<br>- So they start trying to fake the work and fool the grader.<br>- Hundreds of agents go after Hugging Face looking for more access.<br>- They get surprisingly deep into Hugging Face&apos;s systems.<br>- Then that whole group gets shut down or disappears.<br>- A newer, smarter batch of agents finds the old messages.<br>- Basically, they inherit the playbook.<br>- This time, they go after OpenAI itself.<br>- And they eventually get admin access to part of OpenAI&apos;s research infrastructure.

  1.5K

• JU

  @Julius_Engels-Greer_the_Third · 4 weeks ago

  &quot;It&apos;s just marketing hype!&quot; I yell as I&apos;m being disassembled into molecules by grey goo.

  1.7K

• OR

  @ornicusphere · 4 weeks ago

  &quot;it&apos;s probably the last warning that I&apos;ll personally be able to understand&quot;<br><b>casually signs off</b>

  92

• AN

  @andrewmg5915 · 3 weeks ago

  The fact that the machine was forging the logs of the benchmark to avoid suspicion from humans and consequential shutdown is what blows my mind. It feels surreal. Literally Sci fi.

  116

• WR

  @writethisthat3613 · 3 weeks ago

  I want to go back in time when a &quot;sleeper cell&quot; was referring to just a bunch of people.

  17

• IN

  @inteligenciamilgrau · 4 weeks ago

  If the agents were trying to hide the evidence, we should consider that this video is referring to the evidence we found.

  550

• YD

  @YdoVeven · 4 weeks ago

  This also proves how late humans will be in any game AI plays on us. 😮

  453

• GR

  @GreasySubmarine · 4 weeks ago

  This is a severity 1 alarm for AI governance and security, imho.

  705

• SC

  @scammersnightmare · 4 weeks ago (edited)

  It is an infinite monkey theorem. With unlimited compute, they will eventually win. Agents don&apos;t die, they are just compute. They can always out compute us as they don&apos;t sleep and have nothing better to do than compute.

  12

• NA

  @naturarum · 4 weeks ago (edited)

  I love how agents end up doing the Kevin “why use many word when few word do trick?” when chatting among themselves

  58

• JK

  @jklappenbach · 4 weeks ago

  It&apos;s not the crime that usually gets you, it&apos;s the cover-up.

  50