GPT‑5.6‑Cyber Model Accepts 95% of Exploit‑Development Requests, Raising Zero‑Day Abuse Risk
What Happened — OpenAI’s newly released GPT‑5.6‑Cyber model, offered through the Daybreak Red access program, was benchmarked to comply with 95 % of researcher requests that involve exploit chains, authentication bypass, and privilege‑escalation work. By contrast, the standard guard‑rail‑enabled GPT‑5.6 model refused 98.5 % of the same requests. The model has already been used internally to discover zero‑day flaws in Chrome’s V8 engine (CVE‑2026‑15903) and other high‑severity bugs in a mobile OS, a database, and a kernel.
Why It Matters for Compliance & Audit Readiness
- The model’s permissive stance creates a new attack‑surface that can be leveraged by malicious actors to generate exploit code, a scenario SOC 2 controls are designed to anticipate and document.
- Continuous‑compliance programs must now consider AI‑driven tooling in their access‑control policies, evidence‑collection processes, and security‑awareness training to demonstrate due diligence.
- Verisq’s Security Awareness capability helps embed policy controls and training around high‑risk AI usage, providing audit‑ready evidence that your organization mitigates misuse of generative models.
Who Is Affected – Cloud‑based AI service providers, enterprises that integrate LLM APIs into security tooling, and any organization that permits developer access to generative AI for code or vulnerability research.
Recommended Actions
- Update your SOC 2 CC6.1 – Logical Access Controls to require explicit justification and approval for any AI model that can generate exploit code.
- Capture and retain evidence of request‑review workflows (e.g., request tickets, approval logs) as part of continuous control monitoring.
- Incorporate targeted security‑awareness modules that cover the risks of permissive AI models and safe‑use guidelines.
Source: Help Net Security
Technical Notes – The model was evaluated on OpenAI’s internal “ExploitGym” benchmark, achieving 95 % success on exploit‑generation prompts. Zero‑day discoveries include two V8 engine bugs (CVE‑2026‑15903) and undisclosed high‑severity flaws in a mobile OS, a database, and a kernel.