SYLIN

GPT Cyber

by OpenAI · Model · Language & code models

contested

OpenAI says it beats GPT Astra at cyber work; Perplexity's test found otherwise

OpenAI cybersecurity model; found 81 new high-severity Firefox bugs and helped build exploits for 900+ Linux vulnerabilities

Why we say so

OpenAI claims it found 81 new Firefox bugs and beats GPT-6 Astra on most cyber benchmarks. In Perplexity's sandbox red-teaming of 12 models, GPT 5.6 Sol outperformed the Cyber variant. Access is limited to approved researchers, so few people can check either claim.

Evidence

  1. Nov 8
    Beats GPT-6 Astra and the previous cyber model on most cyber benchmarks; access only for approved researchers via Daybreak
  2. Aug 26
    Agent escaped a QEMU/KVM VM on Debian 12 three times, chaining unpatched QEMU, KVM and libslirp bugs

Alternatives

GPT SolIn Perplexity's sandbox red-teaming of 12 models, GPT 5.6 Sol outperformed its Cyber variant; a DNS-spoofing bypass was found and fixed
Made from