1300 166 138 Tell us what you need
News / OpenAI

OpenAI opens GPT 5.6 Cyber to vetted security teams

By Monika Tantau • Published 3 months ago
OpenAI opens GPT 5.6 Cyber to vetted security teams

Key takeaways

• OpenAI has released GPT 5.6 Cyber, a model trained to find zero day flaws and build exploit chains
• It answers 95 per cent of sensitive security queries, against 1.5 per cent for standard GPT 5.6 Sol
• Access runs through two vetted tiers, Daybreak Blue for defenders and Daybreak Red for offensive research
• Australia is one of seven jurisdictions with a Trusted Access for Cyber relationship with OpenAI

Most Read

OpenAI has released a model built specifically to find software vulnerabilities and write working exploits, and it will only hand it to organisations that pass a vetting process. GPT 5.6 Cyber went live on 10 August as part of an expanded Daybreak program, the company’s cybersecurity effort, alongside two new access tiers that separate defensive work from offensive research.

The timing is the interesting part. Three days earlier, OpenAI slowed the rollout of Astra, a more capable model it could not clear against the highest cyber risk level in its own safety framework. The company is now doing two things at once, restricting its most powerful system while deliberately loosening the guardrails on a purpose built one for a small, approved group.

What the new model can do

GPT 5.6 Cyber is built on GPT 5.6 Sol and trained to perform better on the work offensive security teams actually do, including finding zero day vulnerabilities and assembling exploit chains. OpenAI’s developer documentation lists it under the Daybreak family and describes it as its most advanced cybersecurity model for authorised vulnerability research and security testing.

The gap between it and the standard model is not subtle. On OpenAI’s Advanced Cybersecurity Completion Rate benchmark, GPT 5.6 Cyber answers 95 per cent of sensitive security queries, covering scenarios such as exploit chain development, authentication bypass and privilege escalation. General purpose GPT 5.6 Sol answers 1.5 per cent of the same set, because its safeguards refuse almost all of them.

This is the second generation of the idea rather than the first. OpenAI shipped a cyber tuned version of GPT 5.5 under Trusted Access for Cyber earlier this year, and Australia’s government was among those able to reach GPT 5.5 through that arrangement for approved defensive work. What has changed is the capability level, and the fact that access is now formally split in two.

How Daybreak Blue and Red differ

Access splits in two. Daybreak Blue gives defenders GPT 5.6 Sol with customised safeguards for work such as malware analysis and incident response. Daybreak Red is the tier that carries GPT 5.6 Cyber, aimed at researchers doing vulnerability discovery, exploit validation and penetration testing.

Neither tier is self serve. Eligibility requires identity verification, account security controls, monitoring and legal declarations covering how the models will be used. From 1 September 2026, hardware security keys become mandatory on every Daybreak account, including those already approved.

OpenAI has also brought in security vendors including Accenture, IBM, CrowdStrike, Cisco and Palo Alto Networks to build the models into their own products. Direct model access stays with the approved partner rather than flowing through to every downstream customer, and that detail determines how most organisations will actually encounter this.

Why Astra stayed locked down

GPT 5.6 Cyber was rated High for cyber capability under OpenAI’s Preparedness Framework rather than Critical, and that distinction is what made a public release possible at all. Astra did not get the same clearance. On 7 August, OpenAI said preliminary evaluations showed advances in cybersecurity and agentic coding strong enough that it could not rule out the Critical level, and it slowed the release, as Help Net Security reported.

Critical, in OpenAI’s own definition, describes a model that can find and develop working zero day exploits across hardened real world systems without human intervention. The response was isolated testing environments, restricted network and tool access, stronger protection of model weights and additional monitoring.

The two decisions are really one decision. Capability that can be bounded by vetting gets shipped under conditions, and capability that cannot be bounded gets held back. That is a clearer signal about where the frontier sits than any benchmark table.

Why Astra stayed locked down

Why Australian businesses should pay attention

Australia is one of seven jurisdictions where OpenAI has established Trusted Access for Cyber relationships, alongside Canada, France, Germany, Japan, South Korea and European Union institutions. The company has been working with government bodies across those countries on cyber testing, evaluation and standards.

For most Australian organisations this will not arrive as a login. It will arrive inside a product they already buy. CrowdStrike, Cisco and Palo Alto Networks all hold substantial Australian customer bases, and the partner structure means the model sits behind their tooling rather than in the hands of the businesses using it.

Volume is what makes that matter. The Australian Signals Directorate recorded more than 84,700 cybercrime reports in the 2024 to 2025 financial year, roughly one every six minutes, and responded to over 1,200 incidents, an 11 per cent rise on the year before. Ransomware accounted for 11 per cent of those incidents, and 35 per cent of ransomware victims had data posted online, which is where Privacy Act obligations start to bite.

Faster vulnerability discovery cuts both ways. The capability that lets a vendor find a flaw in a client’s stack before an attacker does is the same capability OpenAI is restricting because of what it does in the wrong hands. Australian security teams are better off asking which tier their vendors sit in, and treating model output as something to verify rather than approve on sight, given how often human approval misses threats.

A narrowing window for defenders

OpenAI framed the announcement around a narrowing cyber defence window, and the phrase is doing real work. If models can find vulnerabilities faster than organisations can patch them, the advantage goes to whoever holds access first.

That is the bet behind the vetting: put the capability in defenders’ hands under conditions rather than wait for it to arrive unconditionally from somewhere else. Anthropic made a version of the same argument in June when it opened Mythos to Australian defenders.

For Australian businesses the practical question is not whether to adopt GPT 5.6 Cyber, since almost none will qualify. It is whether the security products they already pay for are about to get materially better, and whether their own patching cycles can keep up if attackers work with comparable tools. The vetting only holds for as long as every capable lab applies one.

Scroll to Top