Sirack's Universe
Kanalga Telegram’da o‘tish
Unfiltered thoughts. Brain dump and hot takes. Mostly about cyber security, malware analysis, reverse engineering, software dev, creative art and tennis (Rafa is the GOAT!) P.S. Opinions are my own; https://ismyapppwned.com https://sandboxbrief.com
Ko'proq ko'rsatishMamlakat belgilanmaganToif belgilanmagan
218
Obunachilar
Ma'lumot yo'q24 soatlar
+17 kunlar
+330 kunlar
Postlar soni
Ma'lumot yuklanmoqda...
Reaktsiyalar
Izohlar
Telegram Yulduzlari
Eng yaxshi postlar bo'yicha
Ma'lumot yuklanmoqda...
Nashrni tahlil qilish
Postlar | Ko'rish dinamikasi | |||||
I feel vendicated :) | 37 | 0 | 0 | 0 | Loading... | |
[This is bonkers reporting] 🤯
Inside Israel’s Secret Operation to Cultivate Ahmadinejad
#weekendread | 57 | 0 | 0 | 0 | Loading... | |
Huggingface's side of the story.
Essentially, China, GLM and open weight are the real winners 😁
https://huggingface.co/blog/security-incident-july-2026 | 88 | 1 | 0 | 0 | Loading... | |
Wub Nat_.mp3 | 89 | 0 | 0 | 2 | Loading... | |
I am a little skeptical about about the chain of events. Hopefully we will get more information down the road.
Couple of questions
1. OpenAI claimed the agent run in a sandboxed environment with no internet access with a singular goal to find solution for Exploit Gym. Okay, what about about the instruction (full prompt) given to the agent?
2. OpenAI claimed the agent used stolen credentials. From where? Why? 😳 Memory, haveibeenpwned or darknet?
3. There were multiple offramp points where the agent could stop "going rogue" but still continued. Either there is an alignment problem or OpenAI deliberately instructioned it 🤷♂
I am not denying the hack took place but don't underestimate the shrewdness of these people. They will never let go a good crisis go to waste 😁
https://openai.com/index/hugging-face-model-evaluation-security-incident/ | 97 | 0 | 0 | 1 | Loading... |

