GPT-5.6 Sol: OpenAI’s Strongest Cybersecurity Model Yet

OpenAI’s GPT-5.6 Sol doubles the exploit-generation success rate of GPT-5.5 on ExploitGym3, turning real-world vulnerabilities into working exploits at 24.9% under the two-hour cap (up from 15.1%), reaching 33.7% with six hours. On ExploitBench2, measuring progression from vulnerable code to arbitrary code execution, Sol scores 73.5% versus GPT-5.5’s 47.9% at comparable token budgets.

Trusted Access for Cyber

Advanced cyber capabilities are gated behind OpenAI Daybreak’s Trusted Access for Cyber program. Verified individuals and organizations gain access to vulnerability triage and validation, malware analysis, detection engineering, and patch validation in authorized environments. Users must enable hardware-backed passkeys by September 1 to retain access to the most cyber-capable frontier models.

SEC-Bench Pro benchmark

On SEC-Bench Pro, which tests proof-of-concept generation on complex software, Sol scores 71.2% versus GPT-5.5’s 45.8% with improved latency. The model supports defensive tasks including secure code review, patching, threat modeling, and blue teaming operations.

Safeguards and dual-use controls

GPT-5.6 ships with OpenAI’s most extensive evaluation period to date, combining human red teaming with large-scale automated testing. The safeguard architecture layers model-level protections with real-time checks, monitoring, and access calibrated by trust level. OpenAI is also restricting access for high-risk entities and in high-risk jurisdictions.

Pricing and availability

The GPT-5.6 family launched on July 9 with Sol, Terra, and Luna variants. On August 30, Luna pricing was cut 80% and Terra 20%, making the cyber-capable models more accessible to security teams operating at scale.

Sources

OpenAI — GPT-5.6 official announcement
Metabase CVE-2026-72898: CVSS 10 SQLi