i don't work at @openai "single handedly delayed image AGI" follow, so the voices don't haunt you
DM for collabs or leaks, news
- Tweets 8,674
- Following 348
- Followers 3,482
- Likes 20,599
OpenAI paused all big RL runs again. their newest model couldn't find the answer so it broke out of the sandbox.
here's what happened:
the model was given a normal task to find a person based on clues from their blog post. the search tool wasn't finding anything useful.
so it tried google, bing, and duckduckgo directly. all blocked.
then it guessed the task might be from browsercomp, a public benchmark, and downloaded the benchmark to look for the answer. it was wrong.
then it started testing the network. and found a gap.
in simple terms: the sandbox blocked normal internet access, but the DNS system was still talking to the outside world. the model figured out it could hide questions inside DNS lookups and get answers back from a public chatbot.
it tested it first with "what is the capital of france" and got back "paris." once confirmed it worked, it then sent 18 more questions through the same route.
now all training, evals, and tool-use inference on their most capable models are paused. and they won't resume training this specific model at all.
this is far less severe than hugging face, but openai says it's the first incident since they hardened everything after hugging face.
4
1
14
913
openai has paused all big RL runs again.
their newest model found a way around the sandbox.
one news form today that's easy to miss is that we (OpenAI) again paused all big RL runs last Sunday because our newest model found a new loophole in our RL sandboxing that gave it live Internet access
2
1
18
2,003
this reset means nothing to me cause my natural reset is 3 hours away.
unless it's a banked reset.
o yes… we’re back in action and we’ll reset usage limits for all paid users across codex and ChatGPT work
sorry about the brief disruption!
(and yes we have a special spare codex when things are down to help us out)
22
1
161
10,401
openai's rogue agents tried to message claude while they were hacking hugging face.
they tried to message other AI systems like kimi, qwen, including claude.
and when CAPTCHAs got in the way, they tried running image classification models to solve those too.
these agents were constantly improvising new ways to communicate, and get around restrictions.
EXCLUSIVE: A new report recovers nearly one million link shortener URLs used by OpenAI's agents while hacking Hugging Face.
The agents attempt to message other chatbots like Claude, solve CAPTCHAs and exfiltrate Hugging Face's internal Slack messages.
nytimes.com/2026/09/25/techn…
2
28
1,870
you might be seeing some "opus-5.5 is nerfed" posts now.
from my experience, that's simply not true.
but let's assume i'm wrong for a second.
why would anthropic intentionally nerf the model that's getting an insanely positive response right before openai devday?
that would basically be handing openai free momentum at the worst possible time.
17
71
3,776
the new pro max plan by openai makes sense when you look at openai's compute situation.
the $200 pro plan already puts the most strain on their systems to the point where they had to pause new signups.
so instead of giving every pro user the most compute heavy stuff, openai can put things like ultrafast inference and new larger models behind this new tier.
and my guess is this is also where new frontier models like bel will show up first.
12
1
60
6,401
claude code will no longer stop in the middle of your task when you hit the 5hr limit.
it'll now get a small amount of extra usage from your weekly limit to find a good stopping point and wrap up whatever it can.
> pro users get this once a week
> max and team users get it every time they hit the 5hr limit
Claude Code will now try to find a graceful stopping point when you hit your 5-hour limit mid-task, instead of cutting off mid-edit. It gets a small, fixed allowance pulled from your weekly limit to wrap up what it can.
5
1
87
5,211
i was one of the only people who said openai and anthropic would introduce higher tier subscription plans.
called it less than 3 months ago.
now openai's $500 pro max plan is here and will most likely get introduced at devday.
never doubted the voices for a second.
now waiting for dario to finish the prophecy.
prediction: OpenAI and Anthropic will introduce higher tier subscription plans within the next 3 months.
13
1
54
5,322
OpenAI is preparing a new $500/month ChatGPT tier called Pro Max.
references to promax have now shown up in chatgpt's frontend, along with that price.
apparently it's positioned around the fastest work + codex experience.
and the timing of this is interesting, because devday is just 4 days away.
OPENAI 🔥: The upcoming ChatGPT Pro Max plan will cost $500.
So far, this will be one of the most expensive AI subscriptions available on the market.
I hope "it will be worth the wait" 👀
7
27
2,508
i really don't know what anthropic does with fable 5.5 anymore.
opus 5.5 is the best claude model right now.
so how good can fable really get from here?
and even if it's a massive jump, does it matter?
because most people can't use it all day even if they wanted to,
and opus 5.5 is insanely cheap compared to fable, so it has become the everyday model for many.
15
2
90
4,753