GPT-5.6-Cyber: OpenAI's Exploit-Permissive AI Model Explained

OpenAI on Monday unveiled GPT-5.6-Cyber, a cybersecurity-focused variant of its GPT-5.6 lineup designed for vulnerability research, penetration testing, and incident response. Built on the GPT-5.6 Sol foundation, the model is purpose-trained to reduce refusals on higher-risk, dual-use cyber tasks such as zero-day discovery and exploit-chain development. Access is gated through a new offering called Daybreak Red, which provides vetted firms with restricted entry to OpenAI's security-oriented models for authorized exploit validation and security testing. The launch builds on GPT-5.5-Cyber, which shipped in June 2026.

To quantify the model's reduced refusal rate, OpenAI introduced an internal benchmark called the Advanced Cybersecurity Completion Rate. The evaluation measures how often models comply with prompts tied to exploit-chain construction, authentication bypass, and privilege escalation. GPT-5.6-Cyber completed 95.0% of these requests, compared to just 1.5% for the standard GPT-5.6 Sol, 2.0% under Daybreak Blue access, and 57.3% for its GPT-5.5-Cyber predecessor. The model also outperformed earlier iterations on the ExploitGym benchmark, though OpenAI noted it sometimes produces shorter vulnerability reports than GPT-5.6 Sol, which slightly weakens its performance on open-ended repository audits and proof-of-concept submissions.

Among its real-world findings, GPT-5.6-Cyber identified CVE-2026-15903, an out-of-bounds read and write flaw in the V8 JavaScript engine carrying a CVSS score of 8.8. The vulnerability could allow a remote attacker to execute arbitrary code inside the V8 sandbox via a crafted HTML page, and the model also discovered a second previously unknown bug that can be chained with CVE-2026-15903 to escape the V8 heap sandbox entirely. Google patched the V8 flaw in mid-July 2026. OpenAI confirmed the same model has since flagged at least five vulnerabilities in a popular mobile operating system, underscoring both the offensive power and the dual-use risk of an AI tuned to bypass the safety guardrails typical of consumer-facing models. Researchers auditing their own environments can run complementary checks with a SSL/TLS checker, a port scanner, and a WHOIS lookup to validate attack surfaces that an AI assistant might help enumerate.