OpenAI announced GPT-5.6-Cyber on 10 August 2026, a variant of its GPT-5.6 Sol flagship trained for cybersecurity work and released only through a vetted tier called Daybreak Red.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source 1 TechCrunch 2026-08-10 GPT-5.6 Cyber built from GPT-5.6 Sol, available only to trusted customer partners at the Red tier; Daybreak launched earlier in 2026 and now split into two tiers; Blue covers incident response, malware analysis and patch validation and is the recommended starting point for most defenders; Red covers purpose trained models for vulnerability research, exploit validation and security testing; approved partners reportedly include Accenture, IBM, CrowdStrike and Cloudflare; OpenAI on defenders having a narrowing window to prepare. Open source In OpenAI's own evaluation the model completed 95 percent of sensitive cybersecurity requests, against 1.5 percent for the generally available version of the same base model, and it was used to find CVE-2026-15903, a high severity flaw in Chrome's V8 JavaScript engine that Google has since patched.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source The stake is that a lab has now published, in one table, the exact size of the capability its refusals were holding back. We assess with high confidence that this release marks the point where frontier offensive cyber capability stopped being a question of model training and became a question of access control, and with moderate confidence that identity verification will prove a weaker gate than the safeguards it replaced.

The number that reframes the story

The four completion rates are the most informative disclosure in the announcement. GPT-5.6 Sol as generally available answers 1.5 percent of the sensitive request set. The same model inside Daybreak Blue, with its system level screening relaxed for approved defenders, answers 2.0 percent. GPT-5.5-Cyber, the prior purpose trained model, answers 57.3 percent. GPT-5.6-Cyber answers 95 percent.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source Two facts follow, and they point in opposite directions.

The first is that the gap between 1.5 and 95 is not a safety filter, it is training. Moving a defender from the public model into Blue buys half a percentage point.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source That is a striking result on its own: the tier that OpenAI calls the recommended starting point for most defenders is, on this benchmark, barely distinguishable from what anyone can already buy.1 TechCrunch 2026-08-10 GPT-5.6 Cyber built from GPT-5.6 Sol, available only to trusted customer partners at the Red tier; Daybreak launched earlier in 2026 and now split into two tiers; Blue covers incident response, malware analysis and patch validation and is the recommended starting point for most defenders; Red covers purpose trained models for vulnerability research, exploit validation and security testing; approved partners reportedly include Accenture, IBM, CrowdStrike and Cloudflare; OpenAI on defenders having a narrowing window to prepare. Open source 2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source Whatever value Blue delivers, it is not measured by this number. The capability lives in the Cyber variant, which is a separately trained artifact, not a permission flag.

The second is the slope. GPT-5.5-Cyber to GPT-5.6-Cyber is 57.3 to 95 in a single generation.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source Whatever the benchmark measures, and OpenAI's description of it as exploit chain development, authentication bypass and privilege escalation scenarios is the company's own characterization rather than an independently audited suite, a 38 point gain in one release cycle is the relevant trend line for anyone modeling how long the defensive advantage lasts.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source

Access as the only remaining control

OpenAI's answer to the obvious objection is structural. Red is gated: it requires identity verification, legal attestations and approved use cases, and reporting names Accenture, IBM, CrowdStrike and Cloudflare among approved partners.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source 1 TechCrunch 2026-08-10 GPT-5.6 Cyber built from GPT-5.6 Sol, available only to trusted customer partners at the Red tier; Daybreak launched earlier in 2026 and now split into two tiers; Blue covers incident response, malware analysis and patch validation and is the recommended starting point for most defenders; Red covers purpose trained models for vulnerability research, exploit validation and security testing; approved partners reportedly include Accenture, IBM, CrowdStrike and Cloudflare; OpenAI on defenders having a narrowing window to prepare. Open source The CVE disclosure is the proof of concept for the defensive framing: a real vulnerability in a browser engine used by billions, found by the model, reported through coordinated disclosure, patched by the vendor.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source That is precisely the outcome the program is designed to produce, and it is a genuine one.

The difficulty is that access control and capability control fail differently. A refusal trained into weights degrades gracefully: a jailbreak recovers some fraction of the capability, unevenly. A vetted API is binary. It holds completely until an account is compromised, an insider abuses it, a partner's own credentials leak, or a national authority compels access, and then it does not hold at all. Alex Goller of Illumio made the adjacent point from the defender's side, that model guardrails were never the control plane for defense and that controls belong in infrastructure under zero trust assumptions.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source The same logic cuts against the vendor: if guardrails were never the real control, the verification desk that replaced them is now carrying the entire load.

The timing is not incidental. Six days earlier the White House briefed the major labs on a voluntary framework giving the government up to 30 days of pre release access to frontier models, with offensive cyber capability as a central evaluation concern and the framework text kept confidential among participants.3 Fortune 2026-08-04 White House briefing of leading AI companies on 4 August 2026 on a voluntary evaluation framework giving the government up to 30 days of pre release model access, with offensive cyber capability central and the framework details kept confidential among participants. Open source A lab shipping an explicitly offensive capable model into a tiered access program the week after that briefing is, we assess with moderate confidence, demonstrating a governance answer rather than testing one: the argument being made is that capability plus vetting is a policy the government can live with, and that the alternative, withholding the capability from defenders while attackers train their own, is worse.1 TechCrunch 2026-08-10 GPT-5.6 Cyber built from GPT-5.6 Sol, available only to trusted customer partners at the Red tier; Daybreak launched earlier in 2026 and now split into two tiers; Blue covers incident response, malware analysis and patch validation and is the recommended starting point for most defenders; Red covers purpose trained models for vulnerability research, exploit validation and security testing; approved partners reportedly include Accenture, IBM, CrowdStrike and Cloudflare; OpenAI on defenders having a narrowing window to prepare. Open source

Who gains and who loses

The named partners gain most. Accenture, IBM, CrowdStrike and Cloudflare acquire a capability their competitors cannot buy at any price, which converts a model access decision into a competitive moat in security services.1 TechCrunch 2026-08-10 GPT-5.6 Cyber built from GPT-5.6 Sol, available only to trusted customer partners at the Red tier; Daybreak launched earlier in 2026 and now split into two tiers; Blue covers incident response, malware analysis and patch validation and is the recommended starting point for most defenders; Red covers purpose trained models for vulnerability research, exploit validation and security testing; approved partners reportedly include Accenture, IBM, CrowdStrike and Cloudflare; OpenAI on defenders having a narrowing window to prepare. Open source Large enterprises that buy from those firms gain indirectly. OpenAI gains a defensible commercial position in the one AI application where customers are least price sensitive, and a public credential in the form of the V8 finding.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source

The losers are the organizations on the wrong side of the vetting line. Independent researchers, small security shops, open source maintainers and public sector defenders in smaller jurisdictions face an attacker population that is increasingly capable while the strongest defensive tool is available only to firms large enough to pass a partner review.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source Software vendors face a second order squeeze: if approved partners can generate exploit chains at this rate, disclosure volume against their products rises regardless of who is asking, and patch cycles were already the bottleneck. And every organization running V8 derived code just learned that the discovery cost of a high severity browser engine bug has fallen to a model query by an approved account.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source

The counter-case

The strongest argument for the program is the one OpenAI makes: defenders have a narrowing window to prepare, and a world where only unconstrained actors hold this capability is strictly worse than one where vetted defenders hold it too.1 TechCrunch 2026-08-10 GPT-5.6 Cyber built from GPT-5.6 Sol, available only to trusted customer partners at the Red tier; Daybreak launched earlier in 2026 and now split into two tiers; Blue covers incident response, malware analysis and patch validation and is the recommended starting point for most defenders; Red covers purpose trained models for vulnerability research, exploit validation and security testing; approved partners reportedly include Accenture, IBM, CrowdStrike and Cloudflare; OpenAI on defenders having a narrowing window to prepare. Open source That argument is not rhetorical. If open weight models converge on similar capability within a year, the marginal risk of a gated American model approaches zero while the defensive benefit is immediate and real, which the CVE demonstrates.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source

The thesis that access control is the weak point fails if verification holds. Watch for the absence of evidence: no partner account compromise, no leaked model access, no misuse incident attributable to Red over the next year would be meaningful support for the tiered model as durable policy. It should also be said plainly that the 95 percent figure is OpenAI's internal evaluation on OpenAI's own request set, unaudited by any third party we can read, and completion rate is not the same as success rate: answering an exploit development prompt is not evidence the resulting exploit works.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source The benchmark measures willingness at least as much as it measures ability.

What to watch

  • A published CVE count from Daybreak Red. One V8 bug is an anecdote.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source If OpenAI or its partners disclose a running tally of vulnerabilities found through the program by mid 2027, the defensive case becomes measurable; silence would suggest the results did not justify the framing.
  • The first misuse incident. Watch for any reported compromise of a Red tier account or attribution of a real attack to program access within twelve months. That event, or its absence, is the whole test of access control as a substitute for capability control.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source
  • Whether Cyber goes through the government window. The voluntary framework offers up to 30 days of pre release federal access with offensive cyber as a focus.3 Fortune 2026-08-04 White House briefing of leading AI companies on 4 August 2026 on a voluntary evaluation framework giving the government up to 30 days of pre release model access, with offensive cyber capability central and the framework details kept confidential among participants. Open source Public confirmation that a future Cyber variant went through it would tie the two systems together; confirmation that it did not would show the framework's scope stops short of the models it was written for.
  • Anthropic and Google match, or decline to. If a competing lab ships a comparably gated offensive security model by early 2027, tiered access is the industry standard. If none does, OpenAI is holding a position the rest of the field decided was not worth taking.1 TechCrunch 2026-08-10 GPT-5.6 Cyber built from GPT-5.6 Sol, available only to trusted customer partners at the Red tier; Daybreak launched earlier in 2026 and now split into two tiers; Blue covers incident response, malware analysis and patch validation and is the recommended starting point for most defenders; Red covers purpose trained models for vulnerability research, exploit validation and security testing; approved partners reportedly include Accenture, IBM, CrowdStrike and Cloudflare; OpenAI on defenders having a narrowing window to prepare. Open source
  • An open weight model reaches the 57 percent line. GPT-5.5-Cyber scored 57.3 percent one generation ago.2 Infosecurity Magazine 2026-08-11 Announced 10 August 2026; internal evaluation completion rates of 95 percent for GPT-5.6-Cyber, 57.3 percent for GPT-5.5-Cyber, 2.0 percent for GPT-5.6 Sol under Daybreak Blue and 1.5 percent for generally available GPT-5.6 Sol; CVE-2026-15903 high severity flaw in Chrome V8 found with the model, disclosed to Google and patched; Blue and Red tier definitions; Red requires identity verification, legal attestations and approved use cases; Alex Goller of Illumio on guardrails never being the control plane for defense. Open source The date an ungated open weight release matches that number is the date the gating argument stops mattering, and it is the single indicator worth tracking above all the others.

The industry spent three years arguing about whether frontier models should be able to write exploits. That question is settled and the answer is on a scoreboard. What remains is a smaller, harder question about who checks the identity documents.