OpenAI’s GPT-5.6-Cyber found a Chrome V8 bug now tracked as CVE-2026-15903

On August 10 OpenAI published “Expanding Daybreak as the Cyber Defense Window Narrows.” The post introduced GPT-5.6-Cyber, said it was the first model to reach High cybersecurity capability under OpenAI’s own framework, and described work against Chrome’s V8 engine that produced two previously unknown bugs. Google assigned CVE-2026-15903 and shipped a fix in Chrome 150.0.7871.128 / .129.

The Chrome finding

OpenAI wrote that V8’s optimizing compiler skipped a safety check when converting values to integers, so undefined values could become an unexpectedly large number. Used as an array index, that number could make the compiler omit a bounds check and allow out-of-bounds read/write inside the sandbox. A second bug, also found by the model, completed a heap-sandbox escape. OpenCVE repeats Google’s wording: a remote attacker could execute arbitrary code inside the sandbox via a crafted HTML page (Chromium severity: High). NVD later listed CISA-ADP CVSS 3.1 of 8.8. Google credited OpenAI Codex Security (handle amyb) on July 6; the Stable Channel notes are dated July 16.

Access rules

Daybreak Red is the GPT-5.6-Cyber tier. Daybreak Blue is GPT-5.6 Sol with some system-level cyber guardrails removed. OpenAI said individual members using Trusted Access for Cyber will need Advanced Account Security with hardware-backed passkeys by September 1, 2026, or fall back to default access. Yubico has a preferred-pricing bundle tied to the program.

• CVE-2026-15903: OOB read/write in V8 before Chrome 150.0.7871.128
• Published on OpenCVE: July 20, 2026
• Hardware keys: mandatory for Daybreak cyber access on September 1

Takeaways

• The bug is patched; the news is that a dedicated cyber model found it
• OpenAI’s own write-up is the primary source, not a leak
• The September 1 key rule is a stated access change, not a suggestion

Source: OpenAI