OpenAI 发布 GPT-5.5-Cyber，在漏洞复现基准上击败 Mythos 5

OpenAI’s new GPT-5.5-Cyber just beat Mythos 5 on CyberGym.

CyberGym measures whether an agent can reproduce known software vulnerabilities, so this is quite a strong signal for defensive vulnerability analysis of models.

OpenAI also launched a major push to use GPT-5.5-Cyber and human security teams to fix open source bugs before AI bug-hunting tools flood maintainers with low-quality reports.

Vulnerability discovery is becoming much easier, so the scarce part is now remediation, which means confirming the bug, proving reachability, writing a fix, testing it, and giving humans enough evidence to merge safely.

OpenAI’s initiative is to use GPT-5.5-Cyber as a defensive security worker inside Codex.

It scans code, checks whether a vulnerability is real and reachable, writes a patch, tests the patch, and gives humans evidence to approve it.

Daybreak is OpenAI’s new cybersecurity initiative to help trusted defenders find, verify, and patch vulnerable software much faster using AI.

The new checkpoint of GPT-5.5-Cyber, are all part of the company's limited “Trusted Access for Cyber” program and do not involve a public release.