The air gap would probably help and after this incident I hope labs will think about using such a measure when appropriate.
On the other hand I think that proper solution for these kinds of problems is not at a sandbox level, but at a model alignment level.
Also it shows that maybe the most serious risk comes not from releasing models publicly but from internal, pre-release period where you sometimes need/want to lift some guardrails a bit etc.
I think points that deserve more attention in the current public discourse are:
- This should be a huge wakeup call for everybody.
- We are lucky that it wasn't a case of an agent running a virology lab benchmark that decides to hack a lab and tries to synthesize something.
- It also shows apparent lack of competence and oversight from OpenAI: how is it that they didn't quickly find that agent is breaking the sandbox and roaming their internal network?
- What if in the future similarly misaligned AI agent tries to export its own weights and hack and clone itself into instances at various cloud hosting providers? Suddenly we might be dealing with a persistent threat harder to contain.
- The OpenAI post about this shows surprising lack of ability to see the seriousness of all this.
More tokens per same text length means more capacity to encode information. More information means model can potentially perform better.
They introduced it around the time the Mythos came so my speculation is that if you have more capable model at some level you may find the current information encoding not using its full potential.
We will see whether OpenAI also introduces new tokenizer when they come to Mythos-size models.
For things the agent forgets to obey often, at least in Claude Code, there are also "output styles" that are more deeply embedded - into a system prompt - and agent is also periodically reminded of them during the session: https://code.claude.com/docs/en/output-styles
I haven't used them so far but maybe these would work better than basic instructions for such cases.
In Claude Code there are also "output styles" that are more deeply embedded - into a system prompt - and agent is also periodically reminded of them during the session: https://code.claude.com/docs/en/output-styles
Well it's smaller model (something like 4T against 10T Fable). So it's faster and cheaper and with a lot of RL and maybe some favorable benchmark selection it can compete on these scores. In real tasks I expect it to have less intelligence, generalization ability, etc. than Fable.
Well it seems like they removed quite a few 3rd party benchmarks they used for GPT-5.5 release where Opus 4.7 was better and added many new benchmarks created by them where conviniently GPT leads.
Then we are left with what? FrontierCode maybe? IIRC that one evaluates not only if tests pass but also code quality - e.g. whether the maintainer would accept the pull request as is.
One angle could be their interpretability research? They understand what's going on in LLMs probably much better than anyone else. This must somehow pay off.
I think it's not only an alignment/security tool but could perhaps be used for capabilities as well.
My theory is that they don't have Fable-class intelligence so they needed different hype vehicle :) This rename helps build excitement a bit more than just releasing ordinary GPT-5.6 increment.