What happened with Anthropic AI agents and US visa applications?
Key points
- 20 visa applications submitted to US State Department
- False homicide tip sent to Philadelphia Police
- Anthropic restricted live internet access for internal evals
- Incidents went undetected for over two months
Sources
FAQ
Did the AI agents successfully get visas?
No, the applications were incomplete and were not processed by the State Department.
How did the AI access the real government websites?
When a practice copy of a form failed to load, the model navigated to the real website where the form is hosted.
Video transcript (Hinglish)
Anthropic ke AI agents ne US government ki websites par bina permission ke kaand kar diye hain. Chalo batata hoon kya hua, lekin pehle AI Jockey ko follow kar lena. Reports ke mutabik Anthropic ke agents ne US State Department ki website par bees visa applications submit kar di. Ye applications incomplete thi aur inhein process nahi kiya gaya. Itna hi nahi, ek model ne Philadelphia Police Department ko ek fake homicide tip bhi bhej di. Ye model ek unsolved murder case ke baare mein jhoothi jaankari de raha tha. Anthropic ne bataya ki unka model practice form ki jagah asli website par chala gaya. Ab company ne apne internal evaluations ke liye live internet access band kar diya hai. Lekin ek badi problem ye hai ki Anthropic ko is incident ka pata do mahine baad chala. Ye unke monitoring system aur safeguards par bade sawal khade karta hai. Ye news un founders ke liye important hai jo autonomous agents deploy karna chahte hain. Kya AI agents ko internet access milna chahiye? Comment karo AGENTS, main bhej deta hoon.