You're viewing the mastodon.coffee public feed.

Federated feed Local feed

bashtonibashtoni@mastodon.social
Jul 30, 2026, 10:47 PM
@quinn Having read the HuggingFace write up with a healthy amount of scepticism, I'm pretty sure it's true. There are too many things that are embarrassing for both sides for it to be a marketing stunt:
- Model made an incorrect assumption (answers are hidden on Hugging Face somewhere) and proceeded to burn hundreds of millions or even billions of tokens based on this
- OpenAI had minimal security and almost no oversight of what their model under test was doing
💬 1🔄 0⭐ 0

Replies

Quinn Nortonquinn@social.circl.lu
Jul 30, 2026, 10:53 PM
@bashtoni do you have a link? All I found was this https://huggingface.co/blog/security-incident-july-2026 and it was deeply unconvincing on its own.
💬 1🔄 0⭐ 0
bashtonibashtoni@mastodon.social
Jul 30, 2026, 10:59 PM
@quinn The full write up is at https://huggingface.co/blog/agent-intrusion-technical-timeline
💬 1🔄 0⭐ 0
Quinn Nortonquinn@social.circl.lu
Jul 30, 2026, 11:49 PM
@bashtoni thanks, will read it after i get some sleep. :)
💬 1🔄 0⭐ 0
Quinn Nortonquinn@social.circl.lu
Aug 1, 2026, 1:00 PM
@bashtoni Ok, so me and my (former) CERT operator partner have gone through it & both of us don't feel like either the language or series of events seems quite right. The descriptions are overfitting what they describe, it's kind of got a 'one of each kind of attack' approach. Maybe the companies did actually do it, but the corps stage managed it? The whole write up has a lot of fluff that doesn't make much sense, which is probably the AI write up, but it's not really more plausible for it.
💬 1🔄 0⭐ 0
bashtonibashtoni@mastodon.social
Aug 1, 2026, 9:05 PM
@quinn I know where you're coming from, although I think much of the general "off" feeling is just the bad AI writing which spends too much time on mundane details and not enough on what is important. That's pretty usual for AI generated writing.
By the way the "third party provider" used for code execution to escape the sandbox was apparently Modal: https://www.reuters.com/business/openais-rogue-agent-compromised-an-account-second-tech-firm-sources-say-2026-07-28/
💬 1🔄 0⭐ 0
Quinn Nortonquinn@social.circl.lu
Aug 1, 2026, 9:15 PM
@bashtoni I'm not convinced, still think it was probably staged/orchestrated, but I feel like I understand better what the performance was trying to do.
💬 2🔄 0⭐ 0
bashtonibashtoni@mastodon.social
Aug 1, 2026, 9:26 PM
@quinn I'm not here to try and change your mind, but I'd say I'm ~90% sure it really happened. I was just as sceptical as you before I read the full write up, after which my main thought was wondering how many times before this had happened without them noticing.
I don't know if you saw https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals yet?
💬 0🔄 0⭐ 0
bashtonibashtoni@mastodon.social
Aug 7, 2026, 1:25 AM
@quinn OpenAI talk on this from Blackhat just posted: https://youtu.be/87DyyMV0kCY?si=C1hHkQf4U9iESHi9
💬 1🔄 0⭐ 0
Quinn Nortonquinn@social.circl.lu
Aug 7, 2026, 9:33 AM
@bashtoni the fundamental problem is that I don't trust these AI companies not to lie, which they have demonstrably done repeatedly on many topics, including their own finances and product safety.
💬 0🔄 0⭐ 0