Conflict & Security
Researchers Breach OpenAI Systems Using Anthropic's Claude AI Models
September 19, 2026 at 04:57 PM
Coverage
3 reports · over 2 hours
يواليوم السابع
باحثون يستخدمون Claude لاختراق OpenAI.. ما النتيجة؟ - اليوم السابع
(opens the publisher’s site)
AJAl Jazeera Arabic
معركة الذكاء الاصطناعي ضد نفسه بدأت.. من ينتصر؟
(opens the publisher’s site)
AJAl Jazeera Arabic
في 3 أيام وبأقل من 3 آلاف دولار.. فريق يخترق أنظمة "أوبن إيه آي"
(opens the publisher’s site)Three Hacktron AI researchers breached OpenAI's internal systems in under 72 hours using Anthropic's Claude models, accessing employee ChatGPT and Codex accounts.
- The three researchers from security firm Hacktron AI accessed the employee accounts and modified an OpenAI repository on GitHub before halting the test.
- The attack, launched July 25, exploited a memory error in the libheif library via HEIF image uploads to OpenAI's community forum.
- Claude Opus 5 produced a working exploit within hours, after its predecessor Opus 4.8 had failed across multiple sessions.
- OpenAI paid the Hacktron team $6,500 under its bug bounty program after the researchers reported the flaws, which cost roughly $3,000 in AI tokens to exploit.