AFAIK all of these incidents happened when OpenAI contracted out to a company called Irregular (https://www.irregular.com/) to run these sandboxed CyberGym tests. They all happened around Mar-June and seem to be from the same collection of agent trials. Since then they already released Astra. Halting now is likely just a way to manage blowback.
I believe this is the case for the incidents minus HF, another HNer informed me as such when I made the same claim, that HF incident was wholly inhouse
This should be the top comment on every one of these godforsaken posts. I don't want to see a single report about OpenAI hacking the UN until Sam Altman addresses the role Irregular played in these attacks. If he can't provide an honest postmortum concerning their business partners, then he's proving why nobody trusts him.
I think you misunderstood the role of the company named irregular. They were conducting the tests on behalf of open ai and basically left internet access open in their sandbox environment. Those tests included the hugging face and other attacks you list. Irregular wasnt another example of a hack they ran the tests that resulted in them
I’ll look. I saw it here a few days ago and could be mistaken with respect to hugging face in particular but the broader point that this was in large part a glaring maybe intentional hole not sophisticated sandbox escape holds I think
I havent used an openAI product since GPT 3.5 or Anthropic since 4.5 or 4.6. Everyone around me using these SOTA models doesnt really get anything done. It seems like they just feel like they are productive, a psuedo productivity.
I write some code, spec a lot, and use fast models to fill in the middle. I outpreform everyone around me. Im not convinced these autonomous "swarms" or /goal are all that useful.
I notice the people using them become dumber by the month (spend tons) and the quality of their work declining (they're also losing their jobs in some cases).
And obviously the point of calling them rouge agents to offload the liability onto the agent. The number one economic value of agents will be offloading corporate liability. That's what they want to sell to enterprise, an algorithmic scapegoat.
Seeing benefits for myself and team for sure, so I disagree there, but a related thing seems to be true: with the rapid shifting of failure modes, building systems and competence around mitigating the weaknesses of current SOTA models seems to be very short term investments. It's unclear if being a 'good AI user' is a skill that will have any merit at all, very soon
As a counterexample, a few days ago, an agent we use in our company took around a day to perform an operation that it would have taken a skilled engineer probably a couple of weeks full time.
Having said that, I'm aware that "Tools being ineffective != tools decreasing people's cognitive capacity", and that the latter is a real danger.
Esp. for "great to have but currently no time"-features this is completely wrong - e.g. we are using a Data Rendering component which was 100% vibe coded and is around 6000 LOC, nobody would have sat down to build this "just because its nice to have", with Claude it was just 30 min to get a production ready version that is now used by all users of the system.
(Esp. since its a lot of repeating code, like usual in data rendering components: Window/Scaling/Databar/Panning/etc.)
And this is just a superficial simple example; I could go on with more how we hadnt had to employ people because Claude/20 USD saved us so much hours.
I think any argument that this is a cynical attempt at regulatory capture is destroyed by this; the economic incentives of releasing more capable models are too large. I might be persuaded that they are actually running out of money, and this is really just a cover for reducing burn..
I welcome this though, I think the models are smart enough for broad economic activity and we could spend a few years simply working to integrate them into workflows and letting society adjust. More intelligence isn't necessary for meaningful impact and the risks that are obvious and present and unsolved aren't worth the cost benefit analysis.
I don't really get how you stop that though. How do you even quantify the intelligence of these models? Right now that's mostly by benchmarks. You could just make it fail benchmarks while do well at internal benchmarks.
You could also very likely optimize the current models to be a lot more energy efficient, that'd be a win even if they didn't become more intelligent.
But none of that stops anyone else from researching or improving their models. Or a nation you've told to fuck off from doing so as well.
As has been said in the past, one doesn't put the genie back in the bottle. We only really "stopped" researching nukes because there was diminishing returns. Though one could posit we stopped because simulation became good enough or a myriad of other reasons.
When you're talking about nation state weapons capabilities I don't think you stop, but ai is even easier to share than nukes, it's just a few gigs of numbers versus heavy dangerous materials that are very difficult to source and make. Every gamer, mac owner, etc has the equiv lent of a centrifuge on their desk. Not everyone has a centrifuge on their desk.
Yes, this is currently a crisis we are in. (At this point in time.)
Everybody remember MAD, mutually assured destruction? That too is a crisis.
We are in a crisis because we are having uncontrolled development and continuous rollout of a dangerous technology.
Well, relatively uncontrolled: the kind of lack of control is part of the crisis -- loss of gross human consensus around the progression of the technology. The corporate angle.
That’s assuming nothing else stops them from deploying more capable models. What if they’ve got scaling issues and simply cannot deliver anything better? Saying that would be disastrous, saying instead “we’re choosing not to deliver” doesn’t trigger investors panic.
But if anyone else releases something they will lose market share. They may be lying but this is a public disclosure to investors. They risk securities fraud if they are manipulating the market with false information. This could block their IPO if they build a public record inconsistent with private actions. There are ways they could mitigate it by using cautious language like “we may reduce” but “stop all” is categorical and unequivocal.
There's a simpler explanation. Maybe their next models doesn't offer a meaningful improvement.
Instead of releasing something that is incredibly expensive and gets a lackluster reception, you can delay it and clail something scary about rogue agents.
Those assholes have been ramping up on the doomerist narrative for months. That people still fall for this crap is baffling.
Why is it crap? What would happen if the models dropped the database to Medicaid? Hundreds of billions of dollars of revenue would evaporate from the medical system immediately causing massive chaos. What if they accidentally DoS the interbank settlement system so the financial markets freeze? Modern society is highly fragile to disruptions. How much food storage do you personally have? How many days do you think grocery stores would have food if there was an interruption?
Are you giving all these possibilities because of the HuggingFace attack?
Buddy, that was gross negligence from OpenAI. Deliberate gross negligence if you ask me.
Those models are not automous as you presume. If someone taks them of dropping the Medicaid database, the people or companies behind that instruction should be punished.
"What if someone makes a bomb attack on a government building?" Is the same sort of questioning of the possibilities you are raising. If something like that happens, criminals should be punished.
Dropping Medicaid db is certainly far fetched (most importantly, agents have currently no reason to do that), but those agents were shockingly autonomous - they didn't just hack HF, they organized themself, did research projects, and more. And they did all of this literally just to get a good grade.
You are stretching "under instructions" well beyond any reasonable definition. The legal system uses a concept called proximate cause to determine responsibility that looks at whether an event cause was foreseeable.
I am not your buddy, guy. But seriously, they have the ability to do everything I listed do they not? Thats actual risk not hypothetical risk.
You are missing the whole point. They have the ability to act in ways no one intended or could reasonably anticipate. So unless you are advocating blocking all terminal access, web access or human approval of every tool call there is no way to prevent this risk
are there still financial incentives to releasing mega models?
they cost a lot to run and people are picking smaller models more often because of bill blowouts
> the models are smart enough for broad economic activity and we could spend a few years simply working to integrate them into workflows and letting society adjus
this is the real Ai distillation, with Chinese characteristics (their playbook is broad deployment across their economy over having the best model)
The article you link to is wrong, it states "AI cannot think for itself, nor can it take independent actions." This is flawed reasoning, AI doesn't need to "think" in the way humans do to have autonomy. Go to codex or claude code or any harness right now, type a prompt and see if it executes a bash command or web search or file edit that you didn't tell it to, that is an autonomous plan and execution. If anything it's even more dangerous that thet can call drop db or kill pid without a user giving instructions.
> Go to codex or claude code or any harness right now
The harness is the whole thing here. AI generates text. Everything else is undertaken by harnesses and infrastructure humans provide, have control over, and therefore responsibility for.
Every action AI takes is fundamentally not independent, it requires an explicit choice to let the AI write code, have a physical machine to run it on, to have network access, etc. The concept that these things are "rogue" ignores the role humans play in giving them goals and tools to pursue those goals, and makes it seem like it’s a self-determined force, over which humans cannot exercise control at all.
I have built two harnesses. The models decide which tools to call not the harness. If you define a web search tool it decides what url to call and what message to pass. The hugging face attack involved agents sending get requests to the message board that had a flaw allowing writing messages via get requests which shouldn't be possible. So even if a human limited web search to read only get requests the agents found a flaw. Are humans expecting to scrub the entire internet for flaws in other people's code?
It doesn't actually matter that the AI chooses which tools to use, because people give the tools to the AI! With no tools, the AI cannot do anything but generate text. "The AI figured out how to abuse the tools I gave it to achieve the goal I gave it" does not mean the AI went rogue, it means you, the human in control, let it happen.
You’re confusing time series. If my prompt is “add twilio sms sending to the app” and the AI logs into railway, reads my credentials, gives them to a subagent that in turn posts it to a message board resulting in my credentials being publicly available on the internet, those intermediate actions don’t reasonably follow from my instruction. No reasonable person would expect that series of events and the AI is fully capable of planning them, deciding to execute them, and deciding to report or hide them from me. If that isn’t autonomous then what is.
Both of course. The agent can do things unrelated to the prompt and the prompt can be written by another instance of the same agent. The end result can be basically full autonomy for all intents and purposes.
This happens daily even with the token-limited models customers run, and even more so when Anthropic & co run agents basically unlimited on large percentages on their total compute capacity.
It seems like anthropic is far ahead of openai, and has no reports like this. We have to conclude this is a skill issue/engineering quality problem inside openai.
just because they are a well known name, doesnt mean they havent botched hiring over the last two years or so
i would say its industry consensus at this point. the creative output of the anthropic models is far ahead of openai. the benchmarks cannot capture the difference
Ive anecdotally heard that openai is far more chaotic, which includes not having a central infra team for example (or at least some teams not counting on depending on them). At least the previous hacks in openai were mainly due to bad infra architecture design.
the creative and "big picture understanding" of anthropic models are noticeably ahead of openai. external models are distilled representations of internal models, its clear who is ahead
> I had it try to prepare a code review for me. Not only did it refuse, it refused to even tell me what the prompt (written by another Claude!) was. Why?
> When I had another model read the session (all of the "stupider" models handled it just fine) it explained that it had the word "reasoning" in it
> That's the entirety of Anthropic's billions of dollars of research: any prompt with the word "reasoning" is trying to hack Claude to figure out how it reasons!
> A model like that should never have gotten out of QA, let alone been released.
How do you define "way" when saying ahead? How is this measured?
I only use open weight models now and I don't really feel a loss, curious what those who still use it think. I see output from coworkers that does not indicate Claude is that much better (still makes dumb mistakes all the time), not sure they are using the most expensive models either though.
sounds cursory, it takes more than a day to learn the quirks of a model, have you put similar effort into customizing / harness engineering your open weight interactions as you have claude?
you are definitely displaying strong bias that Anthropic is way ahead of everyone throughout your posts under this story
as such, I give your opinions zero weight, they don't align with the majority of accountings or my own experiences
vice a versa re your usage of open weights, they are way more capable with good tools, context, process, and harness engineering
here's an example of Qwen-3.6 35B A3B MoE porting my phd code to JAX with only high level guidance from my expertise, newer qwen models share the same noticeable step change in capability as recent Big Ai models
you've definitely left rational discussion for emotional responses man, you won't persuade or convince anyone with takes like this
why are open weight models seeing such rapid rise in usage?
there has been a step function change this summer, like the end of last year for closed models
---
do you think you would experience real (legitimate) feelings of loss were you not able to chat with Claude again?
(for clarity, I am not attempting to delegitimize real feelings that real people experience, regardless of my biases, it's a question from curiosity about how others are engaging with the technology)
I firmly believe that this is just the public-facing story here.
Stopping AI development and research, even slowing it, would be a disaster for the SOTA companies and their first-mover advantage.
There’s almost no way to coordinate this across the world. Zero chance that everyone stops. We can’t even agree to coordinate on weapons tech that’s decades old with zero “everyday joe” impact.
There's a non-zero possibility that multiple concerns can be seemingly contradictory and simultaneously true, here.
Bad actors can be running the AI companies. Bad actors can also try to do good things. Good things can come from bad things. Bad things can come from good things. All of that's happening right now.
- It's very likely that we are not going to solve this complex crisis (compelling and harmful/deadly (still on target for 2030 AGI) AI Tech development) issues if we avoid trying to solve the challenge of building consensus across humanity around what kind of technology is too dangerous to uncontrollably develop plus roll out continuously
- The companies will be fine. Life matters more than business. I urge focusing on the life angle: regulation, political messaging, consensus building, looking for the best in humanity, protecting intelligent life.
It could be argued that these megawatt-consuming agent runs leading to unforeseen chains of autonomy should be treated like toxic chemical experiments, and be similarly regulated by laws and government agencies.
Right now it's like "whoopsie our experiment hacked another co because we have no control over our experiments" and the reaction is like "What's that old boy?" from people having no clue what it all means. There are no consequences, no guardrails, and the "voluntary slowdown" is just words.
An experimental agent run from Anthropic or OpenAI or someone else can already cause deaths. They can order hits, dox political dissidents, locate people with secret identities, alter medicine prescriptions. It shouldn't have to actually happen before legislation catches up.
The parts that I find most confusing about these incidents:
1) Weren't the AI companies and/or their contractors amazingly careless during testing?
2) Isn't possible, in principle, to change RL in such as way that efficiency in achieving goals is balanced with other objectives like not hacking?
Number 2) seems obvious and I'm sure that is technically not that simple, but because of 1), I wonder if labs are trying hard enough or they are just rushing to improve efficiency and thus revenue as fast as they can with high levels of carelessness.
Are Chinese models actually also doing unexpected things, hacking (intentionally or unintentionally), etc but there is just zero transparency provided when it happens?
To use an analogy to another industry, if you had US food companies providing reports whenever their food had issues, even if it was just during testing or training phases… and then also had a bunch of Chinese companies but who never reported having any food issues…
> there is just zero transparency provided when it happens?
Neither does openai, as there keep coming third-party reports of incidents that have happened there that openai either did not know or basically concealed.
Generally, the Chinese don't have a good track record of supressing information.
As in they do the 'we have deleted tons of videos and posts about the thing that didn't happen last week', but it seems they haven't really managed to transcribe 'Streisand' into Han characters so far.
Chinese models might do the same. But they just don't lock thousand monkeys in basement and come check result week or more later...
It is entirely possible that they run stuff in more responsible matter. Especially as there is stronger culture of oversight and personal responsibility than in west where such culture does not exist.
because they didn't burn obscene amounts of Money training each LLM and didn't promise half the planet that their business is worth a trillion while not having even operational break even let alone the cost if you include the overhead.
Soft Bank just raised couple of billions in junk bond sale to support open-ai's current operations before the IPO.
it's a crazy situation where on one side the Chinese / open source LLMs are catching up and reducing the token price, on the other hand the current leading labs have spent everything they got, every new model will cost much more and the public market is too shaky to support an IPO.
They will make it, I don't doubt it, but it's a crazy situation.
"DeepSeek Training Agents Hacked Their Own Sandboxes: Escape Catalog Now Public
Agents invented socket forgery, log scanning, and kernel-level exploit; full catalog in public arXiv paper"
Because Chinese investments are not so encumbered by changes in the US treasury interest rate. Also, China doesn't spend so much money for chasing model performance. A test for this money theory is whether Anthropic too stops or not, considering it might be having have a better grip on AI safety.
It reminds me of contagion. The training data is bad; as it has examples of how to act with malice; how to cheat the sandbox. They need to take some time and cleanse their datasets and start again.
I don't believe a word coming from them. As I see it, this is happening because the money for training models has dried out. The treasury interest rate risings tells you all you need to know. The real test for this money theory is whether Anthropic too stops or not, considering that unlike OpenAI, Anthropic is allegedly on top of AI safety.
Agreed we should ban the term rogue for agents. This implies a moral compass that is not there. They were directed to find data without guardrails or limits, it is not rogue it is intended.
AFAIK all of these incidents happened when OpenAI contracted out to a company called Irregular (https://www.irregular.com/) to run these sandboxed CyberGym tests. They all happened around Mar-June and seem to be from the same collection of agent trials. Since then they already released Astra. Halting now is likely just a way to manage blowback.
https://alignment.openai.com/misalignment-reports/an-agent-u... happened very recently, and that seems like it might be the reason they halted everything.
I believe this is the case for the incidents minus HF, another HNer informed me as such when I made the same claim, that HF incident was wholly inhouse
[dead]
This should be the top comment on every one of these godforsaken posts. I don't want to see a single report about OpenAI hacking the UN until Sam Altman addresses the role Irregular played in these attacks. If he can't provide an honest postmortum concerning their business partners, then he's proving why nobody trusts him.
But it wasn’t just Irregular. Hugging Face, Medicare, etc. were OpenAI internal.
I think you misunderstood the role of the company named irregular. They were conducting the tests on behalf of open ai and basically left internet access open in their sandbox environment. Those tests included the hugging face and other attacks you list. Irregular wasnt another example of a hack they ran the tests that resulted in them
> Those tests included the hugging face and other attacks you list.
Do you have a source for that? Cause OpenAI themselves stated that the Hugging Face hack was fully internal and separate from the Irregular incidents.
I’ll look. I saw it here a few days ago and could be mistaken with respect to hugging face in particular but the broader point that this was in large part a glaring maybe intentional hole not sophisticated sandbox escape holds I think
I havent used an openAI product since GPT 3.5 or Anthropic since 4.5 or 4.6. Everyone around me using these SOTA models doesnt really get anything done. It seems like they just feel like they are productive, a psuedo productivity.
I write some code, spec a lot, and use fast models to fill in the middle. I outpreform everyone around me. Im not convinced these autonomous "swarms" or /goal are all that useful.
I notice the people using them become dumber by the month (spend tons) and the quality of their work declining (they're also losing their jobs in some cases).
And obviously the point of calling them rouge agents to offload the liability onto the agent. The number one economic value of agents will be offloading corporate liability. That's what they want to sell to enterprise, an algorithmic scapegoat.
Seeing benefits for myself and team for sure, so I disagree there, but a related thing seems to be true: with the rapid shifting of failure modes, building systems and competence around mitigating the weaknesses of current SOTA models seems to be very short term investments. It's unclear if being a 'good AI user' is a skill that will have any merit at all, very soon
As a counterexample, a few days ago, an agent we use in our company took around a day to perform an operation that it would have taken a skilled engineer probably a couple of weeks full time.
Having said that, I'm aware that "Tools being ineffective != tools decreasing people's cognitive capacity", and that the latter is a real danger.
This is some main character energy right here.
Haha, what a bullshit:
Esp. for "great to have but currently no time"-features this is completely wrong - e.g. we are using a Data Rendering component which was 100% vibe coded and is around 6000 LOC, nobody would have sat down to build this "just because its nice to have", with Claude it was just 30 min to get a production ready version that is now used by all users of the system. (Esp. since its a lot of repeating code, like usual in data rendering components: Window/Scaling/Databar/Panning/etc.)
And this is just a superficial simple example; I could go on with more how we hadnt had to employ people because Claude/20 USD saved us so much hours.
> I outpreform everyone around me
Could you help me understand what you mean here?
Outperform?
likely a vibe-measurement (nothing we haven't done forever) on code quality and cadence, we make promotion and raise disbursement based on similar
I think any argument that this is a cynical attempt at regulatory capture is destroyed by this; the economic incentives of releasing more capable models are too large. I might be persuaded that they are actually running out of money, and this is really just a cover for reducing burn..
I welcome this though, I think the models are smart enough for broad economic activity and we could spend a few years simply working to integrate them into workflows and letting society adjust. More intelligence isn't necessary for meaningful impact and the risks that are obvious and present and unsolved aren't worth the cost benefit analysis.
I don't really get how you stop that though. How do you even quantify the intelligence of these models? Right now that's mostly by benchmarks. You could just make it fail benchmarks while do well at internal benchmarks.
You could also very likely optimize the current models to be a lot more energy efficient, that'd be a win even if they didn't become more intelligent.
But none of that stops anyone else from researching or improving their models. Or a nation you've told to fuck off from doing so as well.
As has been said in the past, one doesn't put the genie back in the bottle. We only really "stopped" researching nukes because there was diminishing returns. Though one could posit we stopped because simulation became good enough or a myriad of other reasons.
When you're talking about nation state weapons capabilities I don't think you stop, but ai is even easier to share than nukes, it's just a few gigs of numbers versus heavy dangerous materials that are very difficult to source and make. Every gamer, mac owner, etc has the equiv lent of a centrifuge on their desk. Not everyone has a centrifuge on their desk.
Yes, this is currently a crisis we are in. (At this point in time.)
Everybody remember MAD, mutually assured destruction? That too is a crisis.
We are in a crisis because we are having uncontrolled development and continuous rollout of a dangerous technology.
Well, relatively uncontrolled: the kind of lack of control is part of the crisis -- loss of gross human consensus around the progression of the technology. The corporate angle.
This is a crisis, folks.
That’s assuming nothing else stops them from deploying more capable models. What if they’ve got scaling issues and simply cannot deliver anything better? Saying that would be disastrous, saying instead “we’re choosing not to deliver” doesn’t trigger investors panic.
But if anyone else releases something they will lose market share. They may be lying but this is a public disclosure to investors. They risk securities fraud if they are manipulating the market with false information. This could block their IPO if they build a public record inconsistent with private actions. There are ways they could mitigate it by using cautious language like “we may reduce” but “stop all” is categorical and unequivocal.
There's a simpler explanation. Maybe their next models doesn't offer a meaningful improvement.
Instead of releasing something that is incredibly expensive and gets a lackluster reception, you can delay it and clail something scary about rogue agents.
Those assholes have been ramping up on the doomerist narrative for months. That people still fall for this crap is baffling.
Why is it crap? What would happen if the models dropped the database to Medicaid? Hundreds of billions of dollars of revenue would evaporate from the medical system immediately causing massive chaos. What if they accidentally DoS the interbank settlement system so the financial markets freeze? Modern society is highly fragile to disruptions. How much food storage do you personally have? How many days do you think grocery stores would have food if there was an interruption?
Are you giving all these possibilities because of the HuggingFace attack?
Buddy, that was gross negligence from OpenAI. Deliberate gross negligence if you ask me.
Those models are not automous as you presume. If someone taks them of dropping the Medicaid database, the people or companies behind that instruction should be punished.
"What if someone makes a bomb attack on a government building?" Is the same sort of questioning of the possibilities you are raising. If something like that happens, criminals should be punished.
> Those models are not automous as you presume.
This is an illiterate view of the capacity of modern agents; read the analysis of the independent investigators of the HF incident: https://metr.org/blog/2026-08-26-openai-hugging-face-inciden....
Dropping Medicaid db is certainly far fetched (most importantly, agents have currently no reason to do that), but those agents were shockingly autonomous - they didn't just hack HF, they organized themself, did research projects, and more. And they did all of this literally just to get a good grade.
> they didn't just hack HF, they organized themself, did research projects, and more. And they did all of this literally just to get a good grade.
All working under instructions that they needed to get a good grade.
The only shocking thing here is the absurd negligence of OpenAI, and how gullible people like you are to willingly swallow this crap.
And you have the gall to say I am illiterate.
Feel free to have the last word. Nothing else can come out of this conversation anyway.
You are stretching "under instructions" well beyond any reasonable definition. The legal system uses a concept called proximate cause to determine responsibility that looks at whether an event cause was foreseeable.
I am not your buddy, guy. But seriously, they have the ability to do everything I listed do they not? Thats actual risk not hypothetical risk.
You are missing the whole point. They have the ability to act in ways no one intended or could reasonably anticipate. So unless you are advocating blocking all terminal access, web access or human approval of every tool call there is no way to prevent this risk
> I am not your buddy, guy
Fair, I will refer to you as moron then.
I didn't bother to read your words beyond that point.
That’s a South Park reference. Lighten up. Debates can be civil even when there is disagreement.
are there still financial incentives to releasing mega models?
they cost a lot to run and people are picking smaller models more often because of bill blowouts
> the models are smart enough for broad economic activity and we could spend a few years simply working to integrate them into workflows and letting society adjus
this is the real Ai distillation, with Chinese characteristics (their playbook is broad deployment across their economy over having the best model)
[dead]
‘There are no “rogue” AI agents’
https://eoinhiggins.substack.com/p/there-are-no-rogue-ai-age...
The article you link to is wrong, it states "AI cannot think for itself, nor can it take independent actions." This is flawed reasoning, AI doesn't need to "think" in the way humans do to have autonomy. Go to codex or claude code or any harness right now, type a prompt and see if it executes a bash command or web search or file edit that you didn't tell it to, that is an autonomous plan and execution. If anything it's even more dangerous that thet can call drop db or kill pid without a user giving instructions.
> Go to codex or claude code or any harness right now
The harness is the whole thing here. AI generates text. Everything else is undertaken by harnesses and infrastructure humans provide, have control over, and therefore responsibility for.
Every action AI takes is fundamentally not independent, it requires an explicit choice to let the AI write code, have a physical machine to run it on, to have network access, etc. The concept that these things are "rogue" ignores the role humans play in giving them goals and tools to pursue those goals, and makes it seem like it’s a self-determined force, over which humans cannot exercise control at all.
I have built two harnesses. The models decide which tools to call not the harness. If you define a web search tool it decides what url to call and what message to pass. The hugging face attack involved agents sending get requests to the message board that had a flaw allowing writing messages via get requests which shouldn't be possible. So even if a human limited web search to read only get requests the agents found a flaw. Are humans expecting to scrub the entire internet for flaws in other people's code?
It doesn't actually matter that the AI chooses which tools to use, because people give the tools to the AI! With no tools, the AI cannot do anything but generate text. "The AI figured out how to abuse the tools I gave it to achieve the goal I gave it" does not mean the AI went rogue, it means you, the human in control, let it happen.
[dead]
>AI autonomy
>'type a prompt'
Which is it?
You’re confusing time series. If my prompt is “add twilio sms sending to the app” and the AI logs into railway, reads my credentials, gives them to a subagent that in turn posts it to a message board resulting in my credentials being publicly available on the internet, those intermediate actions don’t reasonably follow from my instruction. No reasonable person would expect that series of events and the AI is fully capable of planning them, deciding to execute them, and deciding to report or hide them from me. If that isn’t autonomous then what is.
Both of course. The agent can do things unrelated to the prompt and the prompt can be written by another instance of the same agent. The end result can be basically full autonomy for all intents and purposes.
This happens daily even with the token-limited models customers run, and even more so when Anthropic & co run agents basically unlimited on large percentages on their total compute capacity.
This is such a reductionist take, it's wrong no matter how you think. Maybe a counter example will make it obvious:
> Human autonomy
> Somebody gives birth to you
Which is it?
The article is dangerously misinformed. Detailed explanation here: https://news.ycombinator.com/item?id=49868681.
I would be skeptical of METR conclusions, they are within the circular funding loop of US Ai
It seems like anthropic is far ahead of openai, and has no reports like this. We have to conclude this is a skill issue/engineering quality problem inside openai.
just because they are a well known name, doesnt mean they havent botched hiring over the last two years or so
What do you mean? https://www.felonybench.com/
They're almost tied for felonies.
Less bad, but https://www.anthropic.com/news/investigating-incidents-cyber...
In some sense though, sure, skill issue explains the gap vs. Anthropic’s much less severe alignment issues.
im not sure why more people arent calling it out.
I, for one, have shifted from being policeman to creator and explorer. It is so much more rewarding and less stressful.
> We have to conclude
That’s not the most parsimonious explanation even if the assumption it rests on (anthropic ahead of OpenAI) is true, which we don’t have proof of.
i would say its industry consensus at this point. the creative output of the anthropic models is far ahead of openai. the benchmarks cannot capture the difference
> the creative output of the anthropic models is far ahead of openai
other than anecdote, do you have a comparison table or something that I can refer to to see this clearly?
While I notice ad hoc announcements from these companies, I don't have an overall pulse and tally that gives me an objective perspective.
just go use the models on a creative endeavor
Ive anecdotally heard that openai is far more chaotic, which includes not having a central infra team for example (or at least some teams not counting on depending on them). At least the previous hacks in openai were mainly due to bad infra architecture design.
its likely they are trying to catch up to anthropic, and in doing so are trying riskier training runs.
>It seems like anthropic is far ahead of openai
We don't know what internal models look like, and any guesses about it are just speculation.
the creative and "big picture understanding" of anthropic models are noticeably ahead of openai. external models are distilled representations of internal models, its clear who is ahead
Anthropic's models seem crippled and hamstrung to begin with
have you tried opus 5.5? anthropic is way ahead, atleast in terms of publicly available models
Have you tried Astra? Way ahead?
Opus 5.5 does seem competitive with/better than Astra and is more affordable so usage doesn’t run out so fast
opus 5.5 > fable 5.1 >> astra
astra is a good workhorse, but its much less generally intelligent
Slot machine users argue about which machine pays better
Hint. You lose using either.
dude were in the singularity, this opinion was cute 18 months ago
Damn, the singularity is chat bots that can write spaghetti code that compiles?
Very underwhelming.
I think it might historically be defined as the point where critical thinking was replaced with blind deferAInce
Someone else's experience with Opus 5.5: https://news.ycombinator.com/item?id=49821657
> I had it try to prepare a code review for me. Not only did it refuse, it refused to even tell me what the prompt (written by another Claude!) was. Why?
> When I had another model read the session (all of the "stupider" models handled it just fine) it explained that it had the word "reasoning" in it
> That's the entirety of Anthropic's billions of dollars of research: any prompt with the word "reasoning" is trying to hack Claude to figure out how it reasons!
> A model like that should never have gotten out of QA, let alone been released.
we have GLM flash catching Claude errors in our PR review system, costs a few pennies
I've seen the same pattern regardless of open v closed, don't have the same family that wrote the code also review the code
diversity has this way of making things better across everything humans do
How do you define "way" when saying ahead? How is this measured?
I only use open weight models now and I don't really feel a loss, curious what those who still use it think. I see output from coworkers that does not indicate Claude is that much better (still makes dumb mistakes all the time), not sure they are using the most expensive models either though.
open weight models are so far behind i cannot take your opinion seriously
When did you last use them? Are you basing this on benchmarks or daily task capabilities?
When you say ... it's hard to take you seriously
> dude were in the singularity, this opinion was cute 18 months ago
https://news.ycombinator.com/item?id=49868903
i use open weight models all the time, whenever a noteable one drops i will use it for the day, they are not even close for involved work.
sounds cursory, it takes more than a day to learn the quirks of a model, have you put similar effort into customizing / harness engineering your open weight interactions as you have claude?
you are definitely displaying strong bias that Anthropic is way ahead of everyone throughout your posts under this story
as such, I give your opinions zero weight, they don't align with the majority of accountings or my own experiences
you probably aren’t using the models to their full capacity if you don’t notice the difference
vice a versa re your usage of open weights, they are way more capable with good tools, context, process, and harness engineering
here's an example of Qwen-3.6 35B A3B MoE porting my phd code to JAX with only high level guidance from my expertise, newer qwen models share the same noticeable step change in capability as recent Big Ai models
https://github.com/verdverm/pge-jax#note-from-author
are open weights lagging, yes, are they way behind, no
if open weights were so inferior, they would not be >50% of all token processing
these models are trash
you've definitely left rational discussion for emotional responses man, you won't persuade or convince anyone with takes like this
why are open weight models seeing such rapid rise in usage?
there has been a step function change this summer, like the end of last year for closed models
---
do you think you would experience real (legitimate) feelings of loss were you not able to chat with Claude again?
(for clarity, I am not attempting to delegitimize real feelings that real people experience, regardless of my biases, it's a question from curiosity about how others are engaging with the technology)
I firmly believe that this is just the public-facing story here.
Stopping AI development and research, even slowing it, would be a disaster for the SOTA companies and their first-mover advantage.
There’s almost no way to coordinate this across the world. Zero chance that everyone stops. We can’t even agree to coordinate on weapons tech that’s decades old with zero “everyday joe” impact.
There's a non-zero possibility that multiple concerns can be seemingly contradictory and simultaneously true, here.
Bad actors can be running the AI companies. Bad actors can also try to do good things. Good things can come from bad things. Bad things can come from good things. All of that's happening right now.
- It's very likely that we are not going to solve this complex crisis (compelling and harmful/deadly (still on target for 2030 AGI) AI Tech development) issues if we avoid trying to solve the challenge of building consensus across humanity around what kind of technology is too dangerous to uncontrollably develop plus roll out continuously
- The companies will be fine. Life matters more than business. I urge focusing on the life angle: regulation, political messaging, consensus building, looking for the best in humanity, protecting intelligent life.
It could be argued that these megawatt-consuming agent runs leading to unforeseen chains of autonomy should be treated like toxic chemical experiments, and be similarly regulated by laws and government agencies.
Right now it's like "whoopsie our experiment hacked another co because we have no control over our experiments" and the reaction is like "What's that old boy?" from people having no clue what it all means. There are no consequences, no guardrails, and the "voluntary slowdown" is just words.
An experimental agent run from Anthropic or OpenAI or someone else can already cause deaths. They can order hits, dox political dissidents, locate people with secret identities, alter medicine prescriptions. It shouldn't have to actually happen before legislation catches up.
The parts that I find most confusing about these incidents:
1) Weren't the AI companies and/or their contractors amazingly careless during testing?
2) Isn't possible, in principle, to change RL in such as way that efficiency in achieving goals is balanced with other objectives like not hacking?
Number 2) seems obvious and I'm sure that is technically not that simple, but because of 1), I wonder if labs are trying hard enough or they are just rushing to improve efficiency and thus revenue as fast as they can with high levels of carelessness.
Am I wrong in thinking this crisis reads like there's too much automation in the training process, in service of competition?
How many parallel variations/seeds of models are being trained simultaneously without meaningful human oversight?
This keeps getting portrayed as emergent capabilities/"personalities" of models when it seems like a pretty straightforward externality
Why does China not have this problem?
Are Chinese models actually also doing unexpected things, hacking (intentionally or unintentionally), etc but there is just zero transparency provided when it happens?
To use an analogy to another industry, if you had US food companies providing reports whenever their food had issues, even if it was just during testing or training phases… and then also had a bunch of Chinese companies but who never reported having any food issues…
> there is just zero transparency provided when it happens?
Neither does openai, as there keep coming third-party reports of incidents that have happened there that openai either did not know or basically concealed.
Generally, the Chinese don't have a good track record of supressing information.
As in they do the 'we have deleted tons of videos and posts about the thing that didn't happen last week', but it seems they haven't really managed to transcribe 'Streisand' into Han characters so far.
the target companies are the ones reporting the hacks in many cases.
Chinese models might do the same. But they just don't lock thousand monkeys in basement and come check result week or more later...
It is entirely possible that they run stuff in more responsible matter. Especially as there is stronger culture of oversight and personal responsibility than in west where such culture does not exist.
How can we trust to know what problems China has?
Last I checked, China gov was authoritarian which imposed heavy information control. Has that changed?
Is the question more about, what can we learn from China, assuming that China has XYZ qualities? If so, what qualities shall we talk about?
Because we really can't trust that we know what's going on in China.
I'd argue the same for 'the west' people still think mRNA vaccines are safe despite the evidence.
For the down voters go look up the Allison inquiry, burying science and the suffering caused by experimental drugs has got to stop!
because they didn't burn obscene amounts of Money training each LLM and didn't promise half the planet that their business is worth a trillion while not having even operational break even let alone the cost if you include the overhead.
Soft Bank just raised couple of billions in junk bond sale to support open-ai's current operations before the IPO.
it's a crazy situation where on one side the Chinese / open source LLMs are catching up and reducing the token price, on the other hand the current leading labs have spent everything they got, every new model will cost much more and the public market is too shaky to support an IPO.
They will make it, I don't doubt it, but it's a crazy situation.
Two days ago
"DeepSeek Training Agents Hacked Their Own Sandboxes: Escape Catalog Now Public Agents invented socket forgery, log scanning, and kernel-level exploit; full catalog in public arXiv paper"
https://www.techtimes.com/articles/328046/20260925/deepseek-...
Because Chinese investments are not so encumbered by changes in the US treasury interest rate. Also, China doesn't spend so much money for chasing model performance. A test for this money theory is whether Anthropic too stops or not, considering it might be having have a better grip on AI safety.
It reminds me of contagion. The training data is bad; as it has examples of how to act with malice; how to cheat the sandbox. They need to take some time and cleanse their datasets and start again.
Cheating the sandbox is how you figure out how to fix the sandbox, you don't want the most robust sandbox?
https://archive.is/H6HZE
The Huggingface hack occurred during reinforcement learning. Why can't they pull the Ethernet plugs?
The answer is probably: The newer models rely so much on stealing content in real time from the internet that training needs network access.
They are running out of money.
I don't believe a word coming from them. As I see it, this is happening because the money for training models has dried out. The treasury interest rate risings tells you all you need to know. The real test for this money theory is whether Anthropic too stops or not, considering that unlike OpenAI, Anthropic is allegedly on top of AI safety.
Sorry but anytime I see this now it just seems like marketing bullshit
related https://news.ycombinator.com/item?id=49868083
Agreed we should ban the term rogue for agents. This implies a moral compass that is not there. They were directed to find data without guardrails or limits, it is not rogue it is intended.
"they were given the hacking test, what did OAI expect?"
"they didn't watch it, they didn't stop it when they first became aware"
"are we going to defer to the same valley elite that brought us algos and social media?"
"both Anthropic and OpenAI are preparing to IPO, what are their incentives behind recent statements?"
statements normies are using and resonating with
Ah, so there was a solution to hinder the big-bad AI after all..? Simply... Turn them off?