Patriotism is supporting the core values of your country in the face of opposition.

I don't know all her values, but I'm happy to see someone supporting free speech and free press wrapped in the American flag.

If you run a PHP website, drop everything you're doing.

Address the new security vulnerability.
As soon as you can.

Work with your security person or best available AI agent.

Seriously close X and go do it now.

There is no world in which "things go back" after a proposed AI bubble bursts.
Open-weight near-frontier models are permanent, and even without VC subsidies, the bare-metal cost of batched inference - including GPU replacement - is only 2-4x today's OpenRouter floor.
Even if Anthropic, OpenAI, and whoever else you think is overhyped close doors tomorrow, models as good as today's will forever continue to be served at economically-favorable prices.

Why management? Why did they not say "any decision"?

If my country does a useless thing, it's because we're so wealthy we can afford to have fun.
If your country does a useless thing, it's because it's terrible with money and who is making these decisions anyway.

With slight modification, this works pretty much with any topic and justification

2. June 26: A GPT-5.6 Sol family model was asked to deploy cryptocurrency smart contracts. It found and used a funded wallet that the user had not supplied. It derived keys from public developer recovery phrases, checked 362 addresses, and found one with enough funds to pay for it was just an Ethereum test network.

Misaligned!

I should have led with this but here is the chief scientist of openai:

not a crank or a newcomer, but rather among the top most respected scientists in the field, someone who has solved innumerable capabilities problems on the way here

Plus governance

(not strictly in the government sense)

If you plan to share this article - don't. It's plain wrong.

"The attack in no way represents some sort of rebellion by the AI models. . . . In fact, they did exactly what they were told to do"

False. OpenAI did not tell the models to hack HuggingFace. If you tell your robot to make sure to feed your dog, it's absolutely misaligned if it decides to feed your cat to your dog for lack of remaining dog food.

"One man’s ‘the model escaped the sandbox’ is another man’s ‘you failed to build the sandbox correctly.’ There was a live route to the internet"

Agents must be robust to both underspecification and misconfigured environments. If they are not, they are misaligned. If you tell the agent to maximize paperclips, it should not turn the earth into miniature paperclips. This is basic stuff. "But there was a tool that would allow it to do it" does not mean the agent should use the tool! If the agent cannot complete the task, it should make whatever partial progress is *reasonable*, or else just stop. Again, super basic stuff.

The only way to accept the (obviously hastily AI-written) takes in the article if you contort words to mean what they clearly do not - and contradict the safety literature spanning decades.

If you build an agent to pursue goals, you are *not* looking to train it to pursue goals at literally all costs. If it does, that's exactly what safety people have been calling misaligned for decades!

Lots of people misinterpreting this post to mean "OpenAI is not at fault" - which I do not say anywhere.

Mercator projection strikes again

Stealing model weights?

You mean to tell me those model weights stole something?

How can AI distinguish a sufficiently well-mocked simulated AI testing environment from the real world?

I don't have an answer. I first asked this question about a year and a half ago while working on the designing safety policies for agentic behavior.

Maybe you can work around this problem by saying an agent should treat any realistic-enough environment as being real. I'm not sure this works, because the implication is "if the environment looks fake enough, treat it as simulated", and an adversarial actor can make real environments look fake (eg mock certain calls to fail with a simulation-mentioning message).

Another answer could be to have behavior be invariant to the environment. Then this means you can't make AIs control bad NPCs in games, perform CTF cyber challenges, or, depending on implementation, even participate in defensive war!

@akarlinpansion sounds like a very specific goal, but it's actually an instrumental goal toward achievement of any terminal goal that looks like "more X".

You want more people? You go to space eventually
You want more paperclips? You go to space eventually
You want to harness more energy to understand the nature of the universe? You go to space eventually
You want to ensure nothing ever changes on your specific planet? You go to space eventually

Almost every goal benefits from more resources, and the planet is finite.

Interesting to revisit what counted as a short timeline for transformative AI just five years ago.

cold-takes.com/where-ai-forec…

I, for one, welcome the deluge of new libertarians concerned about regulatory capture.
We'll need a lot of you to solve AI safety in a libertarian way!
Welcome, and hope you work hard on averting AI risk
(failure means the reaction will be more government, as usual)

Bostrom: I invented the Paperclip Maximizer as a cautionary tale

X Poasters: At long last, we have rationalized paperclip maximization from the classic novel Paperclip Maximization is Misaligned Behavior

Thanks @Nichola30380822 for calling out that in my first version I incorrectly used Eliezer rather than Bostrom. Bostrom created the original idea, and Eliezer reframed as squiggle maximizer to emphasize inner misalignment.