OpenAI Expands Daybreak With GPT-5.6-Cyber, 95% Task Completion

OpenAI on August 10, 2026 expanded its Daybreak program and released GPT‑5.6‑Cyber, a purpose-trained cybersecurity model built on GPT‑5.6 Sol. The expansion addresses what OpenAI frames as a narrowing defensive window: threat actors are moving toward fully autonomous offensive AI, and general-purpose models with standard safety guardrails are too restrictive for the legitimate security research needed to counter them.

Two Tiers, One Program

Daybreak now ships as two access levels. Daybreak Blue unlocks GPT‑5.6 Sol for authorized defensive work—vulnerability discovery, secure code review, malware analysis, incident response, and patch validation—by removing system-level guardrails that block legitimate security prompts. Daybreak Red goes further, granting access to GPT‑5.6‑Cyber for exploit validation, authorized vulnerability research, and red teaming. Both tiers require identity verification, account security monitoring, approved-use restrictions, and legal attestations. Beginning September 1, 2026, all individual Daybreak accounts must adopt hardware security keys. Organizations can apply at openai.com/daybreak/partners.

The Completion Rate Gap

The core engineering argument for GPT‑5.6‑Cyber is its Advanced Cybersecurity Completion Rate—an internal evaluation covering exploit-chain development, authentication bypass, and privilege escalation scenarios:

The jump from 57.3% to 95.0% between GPT‑5.5‑Cyber and GPT‑5.6‑Cyber directly addresses persistent complaints from researchers who found the predecessor Daybreak Red tier too restrictive for high-friction tasks. SpecterOps CTO Jared Atkinson noted that the model completed in under a day work that earlier models had not resolved after weeks of intermittent effort.

Performance Tradeoffs

GPT‑5.6‑Cyber is not a strict improvement across every axis. On ExploitGym—which tests whether agents can turn known CVEs into working arbitrary-code-execution exploits in controlled environments—it outperforms both GPT‑5.6 Sol and GPT‑5.5‑Cyber. On the internal Vulnerability Discovery and Report Writing evaluation, however, GPT‑5.6‑Cyber trails GPT‑5.6 Sol, a regression OpenAI attributes to the model producing shorter, less detailed vulnerability reports. On ExploitBench—a harder V8 exploitation task with the sandbox active and a 300-turn agent limit—GPT‑5.6 Sol (Daybreak Blue) solves tasks more token-efficiently and scores higher; extending the turn budget to 600 narrows the gap. OpenAI's footnote confirms GPT‑5.6‑Cyber tends toward higher token usage due to more extensive reasoning—a cost consideration relevant to anyone tracking how agentic AI workloads drive API spend.

Real-World Findings

GPT‑5.6‑Cyber has already been used internally to find two previously unknown, chainable vulnerabilities in V8—Chrome's JavaScript engine—that corrupt memory and escape the V8 heap sandbox. Google fixed the chain under CVE-2026-15903, a high-severity flaw in which V8's optimizing compiler incorrectly skipped a safety check when converting values to integers, allowing undefined values to produce an unexpectedly large number. If that number is used as an array index, the compiler may incorrectly assume it falls within bounds and omit the bounds check, enabling an attacker to read or overwrite adjacent memory. Escaping the heap sandbox required the second vulnerability, which the model also found.

Beyond V8, OpenAI reports using the model to surface at least five vulnerabilities in a popular mobile OS (including a chain from an untrusted app to local privilege escalation), three critical vulnerabilities in a popular database including a remote path to code execution, and over 400 privilege-escalation vulnerabilities in a popular OS kernel. Coordinated disclosure is ongoing.

Safeguards and Preparedness

Under OpenAI's Preparedness Framework, GPT‑5.6‑Cyber reaches the High cybersecurity capability threshold but not the Critical threshold—the same classification as GPT‑5.6 Sol. The model improved over GPT‑5.6 Sol on the specialized tasks it was trained for, but not sufficiently to cross the Critical boundary.

For Codex users inside Daybreak, OpenAI is pushing auto-review mode as the default: it evaluates tool calls requiring elevated permissions before execution and can block actions with significant destructive potential. The security community has already seen the consequences of inadequately scoped AI coding agents—credential leaks through AI coding agent pipelines remain a live attack surface—making auto-review mode less optional than OpenAI's "strongly encouraging" language implies.

The operational gap is now 95.0% versus 1.5% task completion on identical prompts. For security teams still relying on uncredentialed general-purpose models, that differential is their exposure.