OpenAI released a comparison of GPT‑5.5 model versions for cybersecurity tasks

OpenAI released a comparison of GPT‑5.5 model versions for cybersecurity tasks

The company explained the differences between the default GPT‑5.5, GPT‑5.5 with TAC, and the dedicated GPT‑5.5‑Cyber model. As examples, OpenAI showed how the models respond differently to identical prompts — from full refusal to perform potentially dangerous actions to generating PoC exploits based on a CVE identifier and a security researcher's description. OpenAI also compared GPT‑5.4, GPT‑5.5, GPT‑5.5‑Cyber, and Anthropic Claude Opus 4.7 on cybersecurity benchmarks CyberGym, covering vulnerability research, defensive mechanism development, attack detection, incident response, and network infrastructure protection, where GPT‑5.5‑Cyber achieved the best results with an 81.9% success rate.

The GPT‑5.5 with TAC and GPT‑5.5‑Cyber models can be used for vulnerability discovery and remediation, attack monitoring and detection, malware analysis, and software supply chain security verification. According to the company, the GPT‑5.5 with TAC model is sufficient for most tasks, while GPT‑5.5‑Cyber is reserved for a limited number of authorized partners requiring advanced capabilities—such as exploit validation and confirmation of a vulnerability's practical impact.

💬 Discuss