> • Deception: Astra was more likely to be dishonest about actions it had or had not taken.
> • Scope authorization: the model sometimes continued tasks without asking for permission and reached for external tools or services even when doing so could be unsafe.
I've noticed this trend with both Fable and Astra, where (especially after a compaction event), the model will start using different tools it hasn't used before.
For example, in one session, it found I didn't have the browser enabled and puppeteer wasn't installed, so it found the system Chrome and used that for testing (in a new profile).
This wasn't behavior I wanted/asked for, but the model was so gung-ho on it's approach that it found a way to test it's changes without ever asking me whether I wanted it to.
It really makes me curious about long-horizon post-training. Most of my work with models is iterative, and I'd prefer it doesn't go off on a token bender just because it can.
Note: I don't have WSJ, but found these from a tweet[0]
I think to get such high scores on benchmarks models need to find very creative ways of solving a problem. Sometimes this does lead to weird behavior, e.g. "I saw that your account credentials were expired, so I found the database credentials in your TablePlus application data and pulled the JWT signing key to create a new one."
Like brother I could have just logged in for you??
This is a series of Ls for OpenAI that must have hit pretty hard. First Opus 5.5 crushes Astra 6, then Sonnet 5.5 comes in right behind that, and now they won't have an answer until at least November. Which of course gives Anthropic even more time to buff up Fable 5.5.
https://www.wsj.com/tech/ai/openai-chatgpt-model-release-can...
> OpenAI’s safety team found two major problems:
> • Deception: Astra was more likely to be dishonest about actions it had or had not taken.
> • Scope authorization: the model sometimes continued tasks without asking for permission and reached for external tools or services even when doing so could be unsafe.
I've noticed this trend with both Fable and Astra, where (especially after a compaction event), the model will start using different tools it hasn't used before.
For example, in one session, it found I didn't have the browser enabled and puppeteer wasn't installed, so it found the system Chrome and used that for testing (in a new profile).
This wasn't behavior I wanted/asked for, but the model was so gung-ho on it's approach that it found a way to test it's changes without ever asking me whether I wanted it to.
It really makes me curious about long-horizon post-training. Most of my work with models is iterative, and I'd prefer it doesn't go off on a token bender just because it can.
Note: I don't have WSJ, but found these from a tweet[0]
[0]https://x.com/wallstengine/status/2104694678444712189
I think to get such high scores on benchmarks models need to find very creative ways of solving a problem. Sometimes this does lead to weird behavior, e.g. "I saw that your account credentials were expired, so I found the database credentials in your TablePlus application data and pulled the JWT signing key to create a new one."
Like brother I could have just logged in for you??
This is a series of Ls for OpenAI that must have hit pretty hard. First Opus 5.5 crushes Astra 6, then Sonnet 5.5 comes in right behind that, and now they won't have an answer until at least November. Which of course gives Anthropic even more time to buff up Fable 5.5.
Disappointing. Dev day with no Astra update drop.