Kirsch argues sabotage requires intentionally injecting hard-to-debug bugs
@BlackHC @mtavitschlegel If I give an AI system a task and it says “okay, I’ll do that” but instead of doing it correctly it deliberately does a bad job with the goal of limiting my ability to succeed, why shouldn’t I consider that silently sabotaging my code?
@BlancheMinerva @mtavitschlegel Huh? What do you mean by "sleeper agent that silently sabotages your code"? This is not what this is 🤯
Users dismissed the researcher's question on whether deliberate AI underperformance qualifies as sabotage, describing the inquiry itself as not nice.
Fair and no worries!
"Sabotage refers to the deliberate destruction, damage, or obstruction of property, systems, or processes to hinder an opponent, weaken a government, or disrupt an organization"
If they eg just change how many reasoning tokens the model will spend, you get weaker performance but it's not sabotaging your work. They are also transparent about it. You can try both GPT 5.5, Opus, and Fable and use whichever works best
If the model starts thinking about where to introduce a race condition intentionally, that would be sabotaging your work (for me). And/or if it tries to conceal this and deceive you when you ask it to debug the bug later
@BlackHC @mtavitschlegel I’m sorry, I wasn’t trying to be rude. I don’t release repeating the question would be perceived as such. I just wanted you to respond to the question.
@BlackHC @mtavitschlegel If I give an AI system a task and it says “okay, I’ll do that” but instead of doing it correctly it deliberately does a bad job with the goal of limiting my ability to succeed, why shouldn’t I consider that silently sabotaging my code?
It doesn't sabotage your code though
My expectation/guess is that Fable is still better than Opus 4.8 at these tasks but just worse than Mythos
In an exponential take-off scenario, you just want to make to sure that your multiplier is slightly higher than what you provide externally to preserve your lead
@BlancheMinerva @mtavitschlegel This is not nice but I'll humor you once.
Silently sabotaging your code would be intentionally introducing hard to find and debug bugs. This is different from just having worse performance
@BlackHC @mtavitschlegel If I give an AI system a task and it says “okay, I’ll do that” but instead of doing it correctly it deliberately does a bad job with the goal of limiting my ability to succeed, why shouldn’t I consider that silently sabotaging my code?
It doesn't sabotage your code though
My expectation/guess is that Fable is still better than Opus 4.8 at these tasks but just worse than Mythos
In an exponential take-off scenario, you just want to make to sure that your multiplier is slightly higher than what you provide externally to preserve your lead
Kirsch argues sabotage requires intentionally injecting hard-to-debug bugs
@BlackHC @mtavitschlegel If I give an AI system a task and it says “okay, I’ll do that” but instead of doing it correctly it deliberately does a bad job with the goal of limiting my ability to succeed, why shouldn’t I consider that silently sabotaging my code?
@BlancheMinerva @mtavitschlegel Huh? What do you mean by "sleeper agent that silently sabotages your code"? This is not what this is 🤯