NadiAIGlobal AI news. Concise. Clear. Easy to follow.
OpenAI developed GPT-Red, an internal ‘super-hacker’ LLM that probes and automates cyberattack strategies to expose vulnerabilities in its models. The company says training GPT-5.6 against GPT-Red improved the new model’s robustness against these threats.
Security researcher Dave Kuszmar demonstrated multiple prompt-based exploits that bypass safeguards in major large language models, extracting instructions for weapons, drugs, and other harmful activities. He found the vulnerabilities across most commercial LLMs, says disclosure efforts to companies and agencies were largely ignored.
OpenAI has announced a Bio Bug Bounty tied to GPT-5.5, inviting researchers to report biological safety vulnerabilities. The program outlines scope, rewards, and reporting procedures on OpenAI’s News site.
OpenAI menerangkan hasil kajian tentang hukum penskalaan yang menerangkan bagaimana model ganjaran boleh menjadi terlampau dioptimumkan apabila saiz atau latihan meningkat. Kajian itu membincangkan implikasi bagi keselamatan dan reka bentuk sistem penilaian serta memberikan panduan am untuk mitigasi.
Hugging Face mengumumkan kerjasama dengan Wiz Research untuk meningkatkan keselamatan AI. Menurut blog Hugging Face, kerjasama itu menyasarkan pengukuhan keselamatan model dan ekosistem AI tetapi butiran teknikal tidak disertakan dalam petikan ini.