Anthropic defended the decision, calling the exploit low-severity
How to show you don't understand how the technology actually works...
“The Admin asked Dario to fix the jailbreak or de-deploy the model. Dario refused.
— In their blog post, Anthropic defended its decision by saying the jailbreak isn’t serious. “
This is crazy. What are we even doing here?
Many users attacked Anthropic's credibility on jailbreaks as hypocritical and self-serving due to past security lapses and regulatory interests, while insulting its defenders as unqualified lobbyists.
@bradrcarson No idea who you are. But of all the things you can fault me on. Not having a technical understanding is pretty low on the list.
More likely story:
The "jailbreak" wasn't super serious (a situation anyone who has ever received bug reports is familiar with), Anthropic thought the demand to halt model was absurd, and Fed Gov used opportunity to punish and humiliate Anthropic for the prior sins of not bending knee.
Anthropic has more credibility on such topics than Washington
Everyone’s actions make complete sense if this is the case.
The assumption that the jailbreak is both real and serious (aka multiturn, not downgrading to 4.8 like most of the public “jailbreaks”) is the ONE factor that makes the rest of the series of events confusing.
If it’s basically the same “jailbreaks” that have been publicized then everything makes sense.
The govt acted rashly in an area it knows nothing about.
Anthropic is stupid to act like it wasn’t serious to the govt, even if it wasn’t serious.
And Amazon got oneshotted and opted to be a snitchy govt simp. I expect them to pay for this move as Anthropic is not an enemy or a competitor but a literal partner. Very very unbecoming.
@bradrcarson Oh. You run Public First, Anthropic's PAC.
Huh ... you may want to stick to fund raising, or whatever it is you do.
@ErikVoorhees @DavidSacks Punishing a company because a partner found a bypass creates a terrible incentive structure. Frontier labs will simply stop doing external red teaming if it risks triggering a federal shutdown.
@ErikVoorhees @DavidSacks Why not just verify that users aren’t foreign nationals and give them access to the model?
Alternate take: the exploit was probably real but not catastrophic, and the existing Anthropic/government relationship was already strained from poor communication and prior contract friction.
So when the government demanded “fix it or pull the model,” Anthropic may have seen an opportunity to overcomply: take Fable 5 fully offline, generate pressure/press, assert control, and make the mandate look unreasonable.
There’s also probably a cost/sustainability angle. If Fable 5 was extremely token-hungry and expensive to serve broadly, this gave Anthropic a convenient reason to pause or restrict access while blaming the mandate.
My guess: benign-ish jailbreak + rocky relationship + strategic overreaction + cost relief. Not one simple story.
They have no credibility, this is the same company that ever forgot to remove credentials from the contractors and Mythos got out the first time. This is the same company that scream, safety, all while being completely unsafe. The only guard rails they put on that thing was classifiers absolutely ridiculous. You can’t scream that is dangerous and a weapon and then your guard rail failed and say not a big deal. They said guard rails that were weak to begin with because again classifiers.
@bradrcarson You are a government policy / lobbyist class person trying to dunk on one of the most accomplished computing infrastructure entrepreneur and operators in the world
@martin_casado @bradrcarson Your h-index is not even 70 - and how many test of time paper awards did you win this year? Just one? No wonder.
@ErikVoorhees @DavidSacks Anthropic screamed about how dangerous Mythos was… and now won’t fix the training wheels on fable so it can’t become what they’ve warned us about???
@ErikVoorhees @DavidSacks Anthropic has technical credibility but they also have a massive vested interest in being regulated in ways which will increase their shareholder value. All LLMs are jailbreakable. If an LLM is too dangerous to exist without guardrails (as they claim) it ought not to exist
@ErikVoorhees @DavidSacks Far more plausible. Still this is all almost certainly an orchestrated attempt to push through a massive regulatory capture regime to cement incumbency status of today's major players and squash independent and free AI.
@martin_casado @bradrcarson nothing worse than “VC..flipping bozo bit.” Ok maybe “oh a VP what do they know” :) :)
@ErikVoorhees @DavidSacks The problem is thinking the federal government understands tech well enough to make these calls.
They don't. They have never. They will never.
They're fucking dinosaurs in the tar pits, and yet we let them make the call for a world they don't understand & won't long survive.
@ErikVoorhees @DavidSacks on point!
@ErikVoorhees @DavidSacks So something this eh… 🤪🫡
@Le_Frontiersman @ErikVoorhees @DavidSacks Genius, how is a web service with no ID checks supposed to know a user’s nationality? GeoIP? You seriously trust that? You don’t even need a VPN; SNAT is enough. Or should they pressure payment processors to hand over users’ personal data? lol
@ErikVoorhees @DavidSacks I'd like to see Sacks explicitly sign his name to the claims that competitors like OpenAI: 1. Have no jailbreaks like the one invoked here 2. Could simply patch them once identified
I posit that these claims are bullshit, which is why Sacks has to resort to merely implying them.
@martin_casado @bradrcarson when i am king, every time a technical person clarifies something like this the recipient’s phone will zap them a little bit upon reading
