Les tests de sécurité de GPT-5.6 révèlent des failles de jailbreak. Cette version conserve les faits vérifiés et replace le sujet dans une lecture produit, infrastructure et gouvernance.
Ce qui s’est passé
Les tests de sécurité de GPT-5.6 révèlent des failles de jailbreak is the current news peg. The verified record centers on these points:
- UK AISI received pre-deployment access to GPT-5.6 Sol for expert cyber tasks
- Sol scored 95.0% ± 9.8% on AISI CTF expert tasks versus 85.0% ± 11.6% for GPT-5.5
- in The Last Ones 32-step range, Sol completed 7 of 10 runs versus 2 of 10 for GPT-5.5
- OpenAI says GPT-5.6 Sol safeguards blocked about 10x more potentially harmful cyber activity
Pourquoi c’est important
Le signal plus large est agentic cyber risk is now about multi-step task completion after inducement, not only one-turn refusal.
À surveiller
Le prochain point de vérification est how model providers combine identity, environment authorization, audit trails and post-release red teaming.
Sources：Fortune, OpenAI GPT-5.6 System Card, Anthropic notes, CocoLoop.