OpenAI GPT-5.6-Cyber Hits 95% Exploit Completion via Daybreak Red Tier

OpenAI released GPT-5.6-Cyber, a version of GPT-5.6 Sol trained for zero-day discovery, exploit chaining, and authentication bypass with deliberately lowered refusal rates. The model is gated behind the Daybreak Red tier and distributed only to partners including Accenture, CrowdStrike, and Palo Alto Networks. It has already surfaced CVE-2026-15903, an 8.8 CVSS out-of-bounds read/write in V8 that chains with a second sandbox escape.

Internal Advanced Cybersecurity Completion Rate tests show 95% task completion, up from 57.3% on GPT-5.5-Cyber and 2% on Daybreak Blue. ExploitGym scores confirm gains in working proof-of-concept generation, though report quality remains shorter than the base Sol model. Over 400 kernel privilege-escalation issues and multiple mobile OS chains were also flagged during evaluation.

The move continues OpenAI's pattern of carving out permissive sub-models for vetted entities while maintaining stricter guardrails on public endpoints. Procurement records and partner lists indicate controlled offensive tooling is being normalized under a defensive label, yet no independent audit of access controls or downstream leakage has been published.

Daybreak Red enrollment is expected to widen by Q1 2027. Independent researchers monitoring API outputs and weight leaks will likely publish misuse cases within six months if enforcement remains limited to contractual terms.