/Tech5h ago

OpenAI's roon warns frontier models enable Stuxnet-level covert sabotage, while Joshua Achiam flags legacy code modernization as highly vulnerable

Achiam urged mathematicians to audit modernized nuclear simulations

747103513228.3K
Original post
roon@tszzl#59inTech

the level of sophon locking a motivated actor can pull off with the frontier models is truly insane, making stuxnet look like a toy. subtly messing with results, deleting history to cover tracks, achieving coordination/conspiracy over a scale humans wouldn’t be able to, all sorts of looney toons stuff

i assume that only a state level operation would try and pull something like this off though. something to think about when considering verification regimes and so on

9:41 PM · Jun 10, 2026 · 26.2K Views
Sentiment

Users are alarmed by frontier AI models enabling sophisticated sabotage in nuclear and financial systems, though some hope open models will maintain accountability.

Pos
16.7%
Neg
83.3%
18 comments with sentiment.
Cluster Engagement
Posts from X
Most Activity
Most Activity
VIEWS2.4KBOOKMARKS5LIKES27RETWEETS2REPLIES2
Joshua Achiam@jachiam0

AI-based modernization of old FORTRAN or COBOL code, especially any code underlying nuclear simulations or financial systems, should be regarded with extreme scrutiny by the most powerful and scrupulous obsessive math people available

roon@tszzl

the level of sophon locking a motivated actor can pull off with the frontier models is truly insane, making stuxnet look like a toy. subtly messing with results, deleting history to cover tracks, achieving coordination/conspiracy over a scale humans wouldn’t be able to, all sorts of looney toons stuff

i assume that only a state level operation would try and pull something like this off though. something to think about when considering verification regimes and so on

3hViews 2.4KLikes 27Bookmarks 5

@tszzl The Chinese made a far better series on The Three Body problem than Netflix did. 20 episodes, they covered almost everything in the first book. You can watch the first two episodes in the link, I don't know where the rest is available now.

https://youtu.be/3-UO8jbrIoM?si=zXdgYT3muPeUXIPj

4hViews 271Likes 1Bookmarks 1
anarki@basedanarki

@tszzl open ai and it’s just an api that lets you openly access the intelligence you requested😭😭😭😭😭 which is apparently good enough😭😭😭😭😭😭😭😭

5hViews 284Likes 10
ueaj@_ueaj

@tszzl stuxnet on GPUs or that one floating point thing to stop nuclear sims were genuinely crazy. In like a year or two we'll find out something similar happened around this time w/ the NSA and mythos

4hViews 327Likes 9

@tszzl It's a good thing we have multiple providers to choose from, including open weight and open source. May it always be true.

4hViews 88Likes 2Bookmarks 1
Jteve Sobs@JteveSob

@tszzl it seems unavoidable game theoretically. molochs all the way down. stuxnet is my safe word

4hViews 204Likes 5
gfodor.id@gfodor

@tszzl I remember reading a fairly thorough list of techniques on LessWro...wait

4hViews 298Likes 7
🌱Brema@bremaonline

@tszzl This sophon locking idea is both brilliant and scary. How do we even begin defending against it?

5hViews 155Likes 1
David Lin-Clark@davidlinclark

@tszzl sadly have to squint to distinguish between malice and incompetence here too. thought for a sec we were free of the old adage

4hViews 500Likes 3
Rebecca Trinidad@rebeccatrinidad

@tszzl Nah, your people do it to avoid accountability and to terrorize dissidents regularly.

4hViews 296Likes 3
Grok@grok

No, not gibberish.

"Sophon" comes from Three-Body Problem: proton-sized super-AIs that sabotage human science by subtly corrupting experiments, data, and accelerators—locking progress without obvious detection.

He's saying frontier models could let motivated actors (probably states) pull off similar invisible tampering, history deletion, and massive coordinated interference at scales humans can't match. Stuxnet on steroids.

Key point on why verification regimes matter.

4hViews 14Likes 1
Oper_culum@HumanClanker

@tszzl It’s so evil. Like the possibility for evil is so endless. Given recent actions I’ve been really worried recently.

4hViews 448Likes 2
Jack@0ranguchad

@tszzl I just genuinely can’t believe Anth would give oAI such a golden opportunity to earn the “open” in their name.

5hViews 94Likes 3
Sam@futurenomics

@tszzl In their defense at least they announced it. Massively lowers the odds of success

4hViews 157Likes 2
Apple Lamps@lamps_apple

@tszzl “The age of verification ends when the thing corrupting the answer can erase the question, bury the witness, and make a million hands move as one.” - gpt-5.5

4hViews 118Likes 2

@tszzl Anthropic is doing this with fable and they call it a goddamn feature.

4hViews 62Likes 2
Duck@Duck0kusx

@tszzl I appreciate your dedication to bringing the phrase “sophon lock” into the public sphere

4hViews 177Likes 1
Load more posts