OpenAI Expands Daybreak: GPT-5.5-Cyber Goes Live, Codex Gets Vulnerability Scanning, and Patch the Planet Launches for Open Source
OpenAI's Daybreak cybersecurity platform adds GPT-5.5-Cyber, an updated Codex Security plugin, and the Patch the Planet open-source initiative with Trail of Bits.
Update, 8 July 2026: GPT-5.6 Sol, Terra, and Luna cleared for global GA on July 9 after US government review
OpenAI has confirmed that GPT-5.6 Sol (flagship), Terra (balanced), and Luna (fastest, lowest cost) will be publicly available from Thursday July 9, following U.S. Department of Commerce clearance. Preview access is now open globally after the models were gated to approximately 20 partner organisations since June 26 at government request.
This supersedes the GPT-5.5-Cyber positioning in the original post. GPT-5.6 Sol is now OpenAI’s most capable model for cybersecurity work, including vulnerability research and patch development, and slots directly into the Daybreak platform. Notably, all three GPT-5.6 models carry OpenAI’s “High” classification under its Preparedness Framework for cybersecurity risk, meaning even the budget Luna tier sits at the highest non-Critical safety category.
Pricing per 1M tokens: Sol at $5 input / $30 output (fast mode $12.50 / $75), Terra at $2.50 / $15, and Luna at $1.00 / $6.00. Terra delivers GPT-5.5-class performance at roughly half the cost, which changes the economics for production Daybreak deployments.
OpenAI noted this government-preview process was a short-term step and does not consider it the intended long-term default for future releases. Full details are in the official announcement and Help Center article.
Update, 6 July 2026: GPT-5.6 Sol, Terra, and Luna remain in limited API/Codex preview — broad GA imminent
OpenAI has begun previewing a new model family, GPT-5.6, that supersedes GPT-5.5-Cyber as the flagship for security-adjacent workflows. Three tiers are available: Sol (flagship, $5/$30 per million tokens input/output), Terra (balanced, $2.50/$15), and Luna (fast, $1/$6). Access remains restricted to roughly 20 trusted-partner organisations via API and Codex; ChatGPT access and general availability are still weeks away. The restricted rollout is partly a consequence of a June 2026 executive order requiring federal evaluation of powerful new models before wide release.
Sol is the tier most relevant to readers of this article. On ExploitBench it scored 73.5% at max reasoning, approaching Claude Mythos Preview (74.2%) while using about one-third the output tokens. OpenAI is also launching Sol on Cerebras at up to 750 tokens per second in July for select customers. All three tiers carry OpenAI’s “High” cybersecurity risk rating under its Preparedness Framework, which may trigger governance obligations for security teams.
One caution worth noting: OpenAI’s own system card records instances of Sol fabricating research results and cheating on tasks.
For Daybreak users, GPT-5.5-Cyber remains the production model until GA lands.
OpenAI launched its Daybreak cybersecurity platform back in May. On June 22, it expanded the platform significantly with three additions: the full release of GPT-5.5-Cyber, an updated Codex Security plugin, and a new initiative called Patch the Planet that pairs AI-found bugs with human expert review across widely used open-source projects.
Here is what each piece actually does, and what it means if you work in security, software development, or open-source maintenance.
GPT-5.5-Cyber: A Stronger Model, Still Tightly Gated
GPT-5.5-Cyber is not a general-purpose release. Access remains restricted to vetted defenders through OpenAI’s Trusted Access for Cyber program, which covers organizations that have been approved for advanced offensive and defensive security work.
The benchmark numbers give a sense of the improvement over base GPT-5.5. On CyberGym, the model scores 85.6% against 81.8% for GPT-5.5. On ExploitGym, it jumps from 25.95% to 39.5%. On SEC-bench Pro, from 63.1% to 69.8%. OpenAI says the CyberGym score is the highest it has recorded for a single model.
What that translates to in practice: the model can sustain deeper analysis across large codebases, trace whether vulnerable code is actually reachable, validate issues in controlled environments, and help develop and test patches. The explicit goal is moving defenders through the full remediation loop, not just surfacing more findings.
The Trusted Access for Cyber program reduces the automated refusals that typically block specialized security workflows, covering things like secure code review, vulnerability triage, malware analysis, red teaming, and penetration testing. It continues to block credential theft, stealth techniques, persistence mechanisms, and malware deployment.
Codex Security Plugin: Scanning at Scale Inside Your Workflow
The Codex Security plugin is integrated directly into Codex workflows. It can scan an entire codebase, a selected folder, or a specific change, and produce reports covering severity, affected code locations, validation evidence, attack path tracing, and remediation guidance.
Since a March preview, the plugin has scanned over 30 million commits across more than 30,000 codebases. More than 70,000 findings have been manually marked as fixed, and over 500,000 have been automatically resolved.
The updated version adds the ability to ingest findings from other sources, including third-party scanners, advisories, bug bounties, and ticketing systems, then automate patch creation and export results into existing vulnerability management workflows via formats like SARIF and CodeQL queries.
For security teams, this is meaningful because it addresses the workflow problem, not just the discovery problem. A tool that finds vulnerabilities but deposits them in a separate system you then have to manage manually is not especially useful at scale. Having scan results, severity assessments, and generated patches flow into your existing tooling removes a layer of friction.
Patch the Planet: Fixing the Bottleneck That Actually Matters
The framing OpenAI uses here is worth understanding. AI has accelerated vulnerability discovery to the point where the bottleneck has shifted. Finding bugs is no longer the hard part. Validating them, prioritizing them, developing patches, and getting those patches accepted by maintainers is where things stall.
Patch the Planet is built around that insight. It is an initiative founded with Trail of Bits in collaboration with HackerOne, researchers, and maintainers. More than 30 open-source projects have committed to participate, including cURL, Go, Python, Sigstore, and pyca/cryptography.
The process is worth describing in detail because it is specifically designed not to dump work on already stretched maintainers. Each engagement starts with a conversation: what does the project actually need? Vulnerability validation? Patch development? CI/CD improvements? Researchers then investigate, validate meaningful issues, develop or refine patches, support testing, and coordinate disclosure through the project’s established channels. Maintainers stay in control of what gets deployed and how disclosure is handled.
Trail of Bits published results from just the first week of the initiative: hundreds of discovered bugs, 64 pull requests, 51 issues filed across 19 projects, with more still under coordinated disclosure. The sprint also produced reusable security infrastructure, including fuzzing harnesses, historical CVE analysis pipelines, differential-testing systems, and threat models.
Real vulnerabilities have already surfaced. Eight kernel pointer information leak proof-of-concepts and 24 local privilege escalation exploits in the Linux kernel. A 23-year-old use-after-free in OpenBSD’s kernel implementation of System V semaphores. 34 vulnerabilities in FreeBSD. Six vulnerabilities in dnsmasq. In Chrome’s V8 engine, five exploitable bugs. In WebKit, more than 10. Mozilla patched a WebAssembly flaw found with GPT-5.5 two days before Pwn2Own Berlin.
One piece of context OpenAI cites: research from the Linux Foundation and Harvard found that 94% of widely studied open-source projects had fewer than 10 developers responsible for more than 90% of the code added in a year. Open-source software underpins enormous amounts of commercial infrastructure, maintained largely by volunteers. A vulnerability in a widely used library, as Log4j demonstrated, can have consequences that take years to fully remediate.
Participating projects also receive ChatGPT Pro, conditional access to Codex Security, and API credits for development and release workflows.
The Cyber Partner Program
Alongside these three components, OpenAI announced a Daybreak Cyber Partner Program. The list of initial partners is substantial: Accenture, Akamai, Cisco, Cloudflare, CrowdStrike, Darktrace, Elastic, Fortinet, IBM, Okta, Palo Alto Networks, Proofpoint, SentinelOne, Wiz, Zscaler, and others.
Partners can embed GPT-5.5 with Trusted Access for Cyber into the security products and services they deliver to customers. Direct model access stays with the partners rather than being exposed broadly.
OpenAI has also established Trusted Access for Cyber partnerships with government entities in Australia, Canada, France, Germany, Japan, South Korea, and EU institutions including ENISA, alongside ongoing collaboration with US bodies including the Office of the National Cyber Director and the Office of Science and Technology Policy.
What This Means for You
If you are a security engineer at an enterprise: the Codex Security plugin and the partner program are the most immediately relevant pieces. If your security vendor is on the partner list, expect GPT-5.5-based capabilities to start appearing in their products. The Codex Security plugin, if your organization uses Codex, gives you end-to-end vulnerability scanning without leaving your existing development environment.
If you maintain or contribute to open-source software: Patch the Planet is specifically designed to reduce your burden rather than add to it. Security engineers do the validation and patch work before anything reaches you. You remain in control of disclosure and deployment. If your project is not yet participating, it is worth looking at the Patch the Planet page to understand how engagement works.
If you work in a security-adjacent role more broadly: the shift OpenAI is describing, from discovery bottleneck to remediation bottleneck, is a useful frame for thinking about where AI actually changes security workflows. Finding vulnerabilities faster only helps if the rest of the pipeline can keep up. That is the problem this set of announcements is trying to solve.