OpenAIのGPT-5.5-Cyber、OpenBSDの23年前の脆弱性を発見。要点は明快だ。OpenAI expanded Daybreak with GPT-5.5-Cyber, a security-specialized model scoring 85.6% on CyberGym versus 81.8% for the base model.
重要なのは次の層だ。Reported findings include an old OpenBSD use-after-free issue, exploitable V8 bugs, WebKit flaws and a Firefox WebAssembly fix before Pwn2Own Berlin.
次に問われるのは実行だ。Patch the Planet adds human review from Trail of Bits and coordination with HackerOne so open-source maintainers are not flooded with raw model output.
参考資料: CocoLoopおよび原文で引用された公開報道