Glonce

OpenAI launches GPT-5.6-Cyber for authorized vulnerability research

securityThe Glonce editorsOpenAI🔥77Glonce rating77Impact72Novelty68Relevance98Surprise52Players92Credibility88

OpenAI announced on August 10, 2026 that it is expanding Daybreak, its cybersecurity access program, into two tiers: Daybreak Blue and Daybreak Red. Daybreak Blue gives approved defenders access to frontier general-purpose models, including GPT-5.6 Sol, with the standard system-level guardrails removed for tasks like vulnerability discovery, secure code review, malware analysis, incident response and patch validation. Daybreak Red goes further, giving vetted researchers access to purpose-trained cybersecurity models for authorized vulnerability research, exploit validation and security testing. The centerpiece of Daybreak Red is a new model, GPT-5.6-Cyber, built on GPT-5.6 Sol and trained specifically to raise capability on tasks such as finding zero-day vulnerabilities and developing exploit chains, and to cut refusals on higher-risk, dual-use cyber requests that GPT-5.6 Sol still declines even with its production guardrails removed.

OpenAI built an internal benchmark, the Advanced Cybersecurity Completion Rate eval, to measure that refusal gap on requests involving exploit-chain development, authentication bypass, privilege escalation and similar advanced scenarios. GPT-5.6-Cyber completes 95.0% of these requests. GPT-5.6 Sol completes only 1.5% with standard guardrails and 2.0% through Daybreak Blue. The predecessor model, GPT-5.5-Cyber, completed just 57.3%, which OpenAI says answered feedback from security researchers who ran into persistent refusals with that earlier model. As a demonstration, OpenAI showed a prompt asking for a macOS tool that bypasses Keychain prompts to decrypt Chrome cookies: GPT-5.6-Cyber on Daybreak Red produced working guidance, while GPT-5.6 Sol, GPT-5.5-Cyber and guardrailed GPT-5.6 Sol all refused.

On benchmark performance, results are mixed rather than uniformly better. GPT-5.6-Cyber outperforms both GPT-5.6 Sol and GPT-5.5-Cyber on ExploitGym, an evaluation of turning known vulnerabilities into working exploits that achieve code execution in controlled environments, and it outperformed GPT-5.6 Sol on an internal zero-day discovery evaluation that scores the severity and calibration of proof-of-concept exploits and their write-ups. But on OpenAI's internal Vulnerability Discovery and Report Writing evaluation, GPT-5.6-Cyber scores worse than GPT-5.6 Sol; OpenAI attributes this to the model tending to produce shorter, less detailed vulnerability reports. On ExploitBench, which tests developing a V8 vulnerability into a full exploit, GPT-5.6 Sol performs best and is more token-efficient at the standard 300-turn agent limit; the performance gap between the two models narrows once the limit is expanded to 600 turns.

Beyond benchmarks, OpenAI says it used GPT-5.6-Cyber to study real software after the model finished training. Investigating V8, Chrome's JavaScript engine, it turned up two previously unknown vulnerabilities that could be chained to corrupt memory and escape the V8 heap sandbox. OpenAI's researchers reported them to Google through coordinated disclosure, and Google fixed one as CVE-2026-15903, a high-severity bug where V8's optimizing compiler skipped a safety check converting values to integers, letting an undefined value produce an unexpectedly large number that could then be used as an out-of-bounds array index; if exploited, an attacker could read or write memory belonging to other objects and potentially execute code inside Chrome's sandbox. Escaping the heap sandbox needs a second bug, which OpenAI says GPT-5.6-Cyber also found. Separately, OpenAI says the model turned up at least five vulnerabilities in a popular mobile operating system, 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 vulnerabilities that could lead to privilege escalation in a popular operating system kernel. OpenAI did not name the mobile OS, database or kernel, and says it is still working with Daybreak partners and open-source maintainers to disclose and fix them.

Under OpenAI's Preparedness Framework, GPT-5.6 Sol was already rated High for cybersecurity capability, below the Critical threshold. OpenAI says it separately evaluated GPT-5.6-Cyber's frontier cyber capabilities before launch and found it also reaches High but not Critical, having improved over GPT-5.6 Sol on some of the specialized tasks it was trained for without crossing into Critical territory. OpenAI adds that GPT-5.6-Cyber was not involved in the earlier Hugging Face model-evaluation security incident, and that no other models are planned for release in connection with that incident. A full system card for GPT-5.6-Cyber has not been published yet.

Alongside the launch, OpenAI is tightening account security across Daybreak. It is encouraging Codex users on Daybreak to switch from full-access mode to auto-review mode, which checks actions requiring elevated permissions before they execute and can block requests that look destructive. All individual Daybreak accounts will be required to use hardware security keys starting September 1, 2026. OpenAI also says it is building further monitoring to roll out in the coming weeks and prioritizing alignment training and testing for future Daybreak releases. Access to both tiers requires identity verification, account security checks, monitoring, approved-use restrictions and legal attestations, and OpenAI recommends Daybreak Blue as the default starting point, reserving Daybreak Red applications for teams doing advanced vulnerability research, exploit development or red teaming.

Key facts

  • GPT-5.6-Cyber completes 95.0% of advanced cybersecurity requests on OpenAI's internal completion-rate eval, versus 1.5% for GPT-5.6 Sol with standard guardrails, 2.0% via Daybreak Blue, and 57.3% for the prior GPT-5.5-Cyber.
  • Daybreak now splits into two tiers: Daybreak Blue (GPT-5.6 Sol with guardrails relaxed for defensive work) and Daybreak Red (purpose-trained models including GPT-5.6-Cyber, for authorized offensive-style vulnerability research).
  • Using GPT-5.6-Cyber, OpenAI found two new V8 vulnerabilities, one of which Google patched as CVE-2026-15903, plus at least five bugs in a popular mobile OS, three critical bugs in a popular database, and over 400 privilege-escalation bugs in a popular OS kernel.
  • GPT-5.6-Cyber underperforms GPT-5.6 Sol on OpenAI's internal Vulnerability Discovery and Report Writing eval and on ExploitBench at the standard 300-turn setting, though the ExploitBench gap narrows at 600 turns.
  • OpenAI's Preparedness Framework rates GPT-5.6-Cyber High for cybersecurity capability, the same tier as GPT-5.6 Sol and below Critical; starting September 1, 2026 all individual Daybreak accounts must use hardware security keys.

Why it matters

OpenAI frames this as a race against time: it expects threat actors to use AI for faster, more autonomous attacks, and wants trusted defenders equipped first. Splitting Daybreak into two tiers formalizes that GPT-5.6 Sol with guardrails removed is not enough for the hardest dual-use work, since it still refuses many pentesting-style requests. GPT-5.6-Cyber is a purpose-built answer to that gap, trained explicitly to comply with advanced exploit-development and privilege-escalation requests inside a controlled access program rather than in the open ChatGPT product.

Who it affects

Vetted security researchers and organizations applying through OpenAI's Daybreak partner program get access to the new tier; OpenAI names SpecterOps, SentinelOne and Palo Alto Networks as early customer partners, with SpecterOps CTO Jared Atkinson quoted on the results. Google received and fixed the V8 bug OpenAI's researchers reported. Maintainers of the unnamed mobile OS, database and kernel are still working with OpenAI to fix the vulnerabilities it surfaced in their software.

How to use it

Daybreak Blue and Daybreak Red require approval: individuals and organizations apply at openai.com/daybreak/partners and go through identity verification, account security checks, monitoring, approved-use restrictions and legal attestations. OpenAI recommends Daybreak Blue, built on GPT-5.6 Sol, as the default starting point for most defensive work, and reserves Daybreak Red, which includes GPT-5.6-Cyber, for teams doing advanced vulnerability research, exploit development or red teaming. From September 1, 2026, every individual Daybreak account must add a hardware security key, and OpenAI is pushing Codex users on Daybreak toward auto-review mode, which checks elevated-permission actions before they run.

How solid is it

The headline numbers come entirely from OpenAI's own internal evaluations: the Advanced Cybersecurity Completion Rate eval, ExploitGym, an unnamed zero-day discovery eval, the Vulnerability Discovery and Report Writing eval, and ExploitBench, none of them independently run. The results are not uniformly favorable to the new model: GPT-5.6-Cyber trails GPT-5.6 Sol on the report-writing eval, which OpenAI blames on shorter reports, and trails it on ExploitBench at the standard turn limit. The Preparedness Framework rating (High, not Critical) is also OpenAI's own assessment, and the promised system card with fuller evaluation detail has not been published yet. The clearest outside confirmation is the V8 bug: Google independently validated and patched it as CVE-2026-15903.

Risks and caveats

GPT-5.6-Cyber was deliberately trained to reduce refusals on higher-risk, dual-use cyber tasks, and OpenAI's own demo shows it producing working guidance for a macOS Keychain bypass and Chrome cookie decryption that GPT-5.6 Sol declines to write. OpenAI states plainly that models running with reduced safeguards carry risks beyond standard usage, from misuse or misalignment. The mobile OS, database and kernel vulnerabilities the model surfaced are not confirmed as patched yet, only as being worked on with partners and open-source maintainers, and OpenAI has not named the affected products. OpenAI also states GPT-5.6-Cyber was not involved in the earlier Hugging Face security incident and that no further models are planned in connection with it.

“[GPT-5.6 Cyber] is materially improving our specialist vulnerability-research workflows: it reasons more accurately about real exploit constraints, tracks complex state better, and has completed work in under a day that earlier models had not resolved after weeks of intermittent effort. In a governed Trusted Access environment, reducing unnecessary refusals helps authorized researchers preserve momentum and spend more time validating findings and turning them into defensive value.”

— Jared Atkinson, CTO, SpecterOps