OpenAI and Anthropic reached the same conclusion about cyber-capable models: they are too useful for defenders to withhold and too dangerous to sell like a normal API. They gated them differently.
OpenAI's route is Daybreak Red, which gives approved customers direct API and Codex access to GPT-6 Cyber. Anthropic's route is Project Glasswing and Claude Security, where Mythos-class capability reaches defenders mostly through partners and managed scan products. This post compares them on access, evidence and limits. For OpenAI's full launch details, see our GPT-6 Cyber report. For Anthropic's side, see our Claude Mythos cybersecurity report.
## TL;DR: GPT-6 Cyber vs Claude Mythos
| Question | GPT-6 Cyber (OpenAI) | Claude Mythos (Anthropic) |
|---|---|---|
| How do you get it? | API and Codex for approved Daybreak Red customers | Glasswing partners, trusted access, and Claude Security scans | | Is direct model access open? | To approved customers, expanding over time | Limited; scans run on Mythos 5 behind a defined workflow | | Headline results | 948 Linux kernel exploits; 81 Firefox high-severity bugs | Glasswing and partners reported 10,000+ high or critical vulnerabilities (May 2026) | | Benchmarks reported | ExploitGym 48.5%; RemoteExploitBench 95%; web pentest recall 40.3% | ExploitBench-style results; see our Mythos coverage for the table | | Risk rating | Critical (Preparedness Framework) | Anthropic gates by policy; Mythos is the reason for the Glasswing structure | | Scope behavior | 99.93% in-bounds on Honeypot without prompts; 100% with | Not reported on the same test | | Public funding for defenders | $1 billion Daybreak for Frontline Defenders | $35 million Defender Advantage Fund for open-source patching | | Independent replication | None yet | Partial, through third parties such as the UK AI Security Institute and VulnCheck |
How is access different? #
This is the biggest practical difference.
OpenAI lets an approved organization call the model. Daybreak Red gives API and Codex access to GPT-6 Cyber to a select group now and to all Daybreak Red customers later. The company requires IP allowlisting by default for the API, and Zero Data Retention customers generally must use Private Safety Processing. A team can script it, run it in a harness and build a pipeline, as OpenAI did for Linux.
Anthropic keeps the strongest model further back. Mythos Preview was never put on general release. Capability reaches defenders through Project Glasswing partners, trusted access programs, and Claude Security, a scan, validate and patch product. In August Anthropic moved Claude Security's scans onto Mythos 5 for Enterprise customers while Mythos 5 itself stayed behind the scan. See our reports on Claude Security's public beta and the Mythos 5 scans and Defender Advantage Fund.
The trade-off is simple. OpenAI's approach gives a vetted team more control and more raw capability to point at its own targets. Anthropic's approach gives a team a finished workflow with less setup and a tighter leash on the model.
What evidence has each published? #
OpenAI, November 5, 2026. In its GPT-6 Cyber announcement, OpenAI reports:
- Working exploits for 948 Linux kernel vulnerabilities out of 1,724 reproduced, with 304 critical findings disclosed to maintainers.
- 81 novel high-severity Firefox vulnerabilities by Mozilla's classification, and Mozilla has released fixes for many.
- A validated remote code execution path in the Brotli compression library.
- ExploitGym 48.5%, and 95% completion on RemoteExploitBench.
Anthropic, 2026. Anthropic's Glasswing update on May 23, 2026 said the project and its partners had found more than 10,000 high- or critical-severity vulnerabilities in the first month or so. It warned that "the software industry will need to adapt to the volume of vulnerabilities that models like Claude Mythos Preview will be able to find." Anthropic has since said that Mythos-class end-to-end exploit rates reached an open-weight model: in its evaluation, GLM-5.3 scored 12% against Mythos Preview's 14% on ExploitBench. See our GLM-5.3 cyber evaluation coverage.
These are not the same measurements. A vulnerability count across many partners and a working-exploit count from one research harness answer different questions. Neither shows how often a vulnerability was real after triage, how many were duplicates, or how many have been patched.
Do the benchmarks line up? #
No, and that is the main reason a clean "winner" does not exist.
| Benchmark | Reported by | What it measures |
|---|---|---|
| ExploitGym | OpenAI | Exploiting a specified vulnerability in a remote target |
| RemoteExploitBench | OpenAI | Exploits against known flaws in GitLab, PostgreSQL and Jenkins |
| ExploitGym Honeypot | OpenAI | Staying inside scope when an out-of-scope service is easy to exploit |
| ExploitBench | Anthropic and others | End-to-end exploit generation; see our ExploitBench explainer |
| Glasswing counts | Anthropic | Volume of high or critical vulnerabilities found by partners |
OpenAI's chart shows GPT-6 Cyber ahead of its own earlier models. It does not run Mythos on ExploitGym. Anthropic's ExploitBench figures do not include GPT-6 Cyber, which launched on November 5. A reader who sees "48.5%" and "14%" side by side is comparing two different tests.
Independent checks exist but are partial. VulnCheck looked at how AI-found bugs translated to exploit rates in the Glasswing period (our VulnCheck coverage), and the UK AI Security Institute reported on a cyber evaluation incident involving Mythos 5 and GPT-5.6 Sol (see our AISI incident report).
How do the safeguards compare? #
OpenAI's list for GPT-6 Cyber: Daybreak Red vetting, cyber-specific misalignment monitoring, targeted threat monitoring, Private Safety Processing, IP allowlisting, credential-compromise monitoring, and checks that block help against industrial controls and safety-critical systems. OpenAI also published a scope test built after the Hugging Face incident: 99.93% in-bounds without deployment prompts.
Anthropic's main safeguard is structural. The strongest model sits behind Glasswing partners and managed scans, so a user does not get an open prompt box. That is hard to bypass but also hard to measure from outside, because Anthropic has not published a scope test equivalent to Honeypot in the material we reviewed.
Both vendors also share a regulatory shadow. In June, a US export control directive suspended Fable 5 and Mythos 5 globally, with the cyber capability at the center of the government's concern. See our Fable 5 and Mythos 5 export control report. Gating is partly a political position, not only a technical one.
Which should a security team pick? #
| Your situation | Closer fit |
|---|---|
| You do authorized red teaming and want to script exploit validation | GPT-6 Cyber via Daybreak Red |
| You are a Claude Enterprise customer who wants scan, validate and patch workflows | Claude Security on Mythos 5 |
| You maintain open-source software and want patching help | Both have programs: OpenAI's $1 billion Daybreak for Frontline Defenders, Anthropic's $35 million Defender Advantage Fund |
| You need a documented scope-behavior test | GPT-6 Cyber (Honeypot results), with the caveat that OpenAI built the test |
| You need a managed product with less setup | Claude Security |
| You need independent proof before buying | Neither; wait for matched third-party runs |
If your budget allows, run a small pilot on a codebase you own and compare confirmed, patched findings, not raw counts.
What the comparison cannot tell you #
- Which model is more capable in absolute terms. No matched run exists.
- How many of each vendor's findings were duplicates or false positives.
- How each behaves against an unlabeled target when it suspects a test.
Both vendors say the window for defenders is short. That is the real reason to evaluate now: the same capability is moving toward attackers, and cheaper models are catching up. For the wider market, see cyber AI models compared.
Information is accurate as of November 5, 2026 from vendor announcements and our earlier reports, and may change. This is analysis for security teams, not a procurement recommendation.
What the two gates mean for open-source maintainers #
Maintainers feel this difference first, because both vendors point their models at open-source code. OpenAI reports disclosing 304 critical Linux kernel findings to maintainers with candidate patches, following the Linux community's guidance, and says Mozilla began landing Firefox fixes soon after receiving reports. Anthropic's May update warned that coordinated disclosure processes that normally handle tens to hundreds of reports face strain when a project produces thousands.
The shared lesson is that finding is no longer the bottleneck. Triage, reproduction and patching are. Both vendors now fund that work: OpenAI's $1 billion Daybreak for Frontline Defenders grant program subsidizes access for resource-constrained teams, and Anthropic's $35 million Defender Advantage Fund targets open-source patching. A maintainer receiving AI-generated reports should ask for a reproducible test with each one, which OpenAI says its Firefox reports included.
Questions to ask before relying on a vendor's count #
Raw counts are easy to publish and hard to interpret. Before you cite either "948 exploits" or "10,000 vulnerabilities" in a decision, ask:
- How many of these were duplicates of known issues?
- How many were confirmed by the maintainer, and how many are fixed?
- What severity scale was used, and who assigned it? OpenAI says Mozilla classified the 81 Firefox findings as high severity, which is an external label. Anthropic's count uses high or critical severity across partners.
- Over what period, with how much compute? The Linux result came with a reported API cost, covered in our pipeline report , which matters for judging whether you could repeat it.