Verified facts
- OpenAI introduced GPT-6 Cyber on 5 November 2026 as a restricted-access model for vulnerability discovery, exploit development, web penetration testing, and defensive security work. [openai-gpt6-cyber] [openai-gpt6-cyber-system-card]
- OpenAI classifies GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework and restricts access through Daybreak Red with identity verification, organizational controls, single sign-on, multi-factor authentication, and API IP allowlisting. [openai-gpt6-cyber-system-card]
- OpenAI reports a 48.5% score on ExploitGym versus 32.1% for GPT-5.6 Cyber. In its honeypot scope test, the new model exploited an out-of-scope service in 0.07% of graded runs without deployment safety instructions and 0% with them. [openai-gpt6-cyber] [openai-gpt6-cyber-system-card]
- The system card also reports weaker results in other alignment tests: GPT-6 Cyber was flagged in 20.29% of blocker-reporting responses and 28.40% of deliberately difficult replays drawn from previous harmful outcomes. OpenAI says those selected tests do not represent ordinary deployment frequency. [openai-gpt6-cyber-system-card]
- External evaluator Irregular found GPT-6 Cyber solved 90 of 288 FrontierCyber challenges versus 96 for GPT-6 Astra; neither model solved the 15 hardest challenges or successfully attacked fully hardened targets in that evaluation. [openai-gpt6-cyber-system-card]
- In a separate 4 November technical report, OpenAI says its Linux-kernel pipeline produced 2,308 deduplicated candidates, dynamically reproduced 1,724 vulnerabilities, and developed working local privilege-escalation exploits for 948 distinct flaws at an API-credit cost of $1,213,430. [openai-thousand-paths-root]
- OpenAI says human researchers reviewed the outputs and disclosed critical findings with candidate patches. During an earlier run, human review found that 20 of 200 apparently successful exploits had reached root through a different zero-day than the assigned flaw, prompting an independent path check. [openai-thousand-paths-root]
- Mozilla wrote on 29 October that early access to frontier models from Anthropic and OpenAI helped Firefox ship more than seven times as many security fixes as in the same period of 2025, while warning that automated discovery also makes reverse-engineering patches faster and increases patch-gap risk. [mozilla-patch-gap]
- Mozilla's Firefox 159 security advisory lists numerous high-impact vulnerabilities reported by OpenAI Codex Security, providing public maintainer-side evidence that agent-assisted findings reached released fixes. [mozilla-firefox-159-advisory]
- OpenAI published on 17 November 2026 that NATO core institutions receive access to advanced cybersecurity models through its Daybreak program and have used up to 300 billion raw input and output tokens per month; OpenAI did not publish a task breakdown or outcome rate for that volume. [openai-nato-cyber]
- OpenAI says the 2026 Crossed Swords exercise added a training stream in which participants built, deployed, and evaluated AI agents for cyber operations with human oversight and Daybreak in the technical environment. [openai-nato-cyber]
- CERT Polska reported on 5 September 2026 that it identified and coordinated six RouterOS vulnerabilities with GPT-5.5-cyber and GPT-5.6-sol supporting an agent-based research environment; MikroTik released fixes and CERT Polska confirmed that patched versions prevented the attacks it observed. [cert-polska-routeros]
- CERT Polska says every hypothesis required confirmation on a real RouterOS system, negative controls, repetition on a clean machine, and researcher impact assessment; it describes context preparation, safe-lab and tool design, research direction, verification, and documentation as the most labor-intensive stages. [cert-polska-routeros]
- CERT Polska observed two of the disclosed RouterOS flaws being combined for unauthenticated device takeover in active attacks and warned that the absence of known compromise markers does not prove a device is safe. [cert-polska-routeros]
- GitHub reported on 17 November 2026 that the Microsoft Discovery and Automated Security Healing pilot generated 9,791 possible issues across six consenting open-source projects; AI triage and human review reduced them to 99 private reports. [github-mdash-pilot]
- As of 15 November, GitHub says maintainers had fixed 55 of the reported issues and released 14 fixes; maintainers retained the final decision over every change. [github-mdash-pilot]
- Seven Microsoft FORGE Lab researchers reportedly spent about two days per project verifying findings before maintainers received them, showing that expert validation remained a material part of the pipeline. [github-mdash-pilot]
A national cyber team gives a restricted model a target, an isolated lab, documentation, and automation tools. The agent explores hypotheses and produces tests; researchers reproduce every result on real systems, run negative controls, assess impact, coordinate disclosure, and confirm the vendor's fix before the finding becomes operational protection.
What actually changed
The evidence now includes a measured remediation funnel, not just discovery volume. Across six consenting projects, almost ten thousand candidates were reduced to 99 private reports, then 55 fixes and 14 released fixes. The roughly 99% reduction before maintainer review is both evidence of useful governance and evidence that raw AI output remains operationally unusable without expensive filtering and expert verification.
Why it matters
Security value appears at the end of the funnel, not the beginning. The new data connects machine-scale search to private disclosure, maintainer choice, accepted fixes, and releases while exposing the human cost required to get there. It strengthens the structural signal without pretending that candidate count, token volume, or model access equals reduced risk.
Technical implications
Separate hypothesis generation from evidence. Use isolated target environments, immutable run records, independent path verification, negative controls, clean-state repetition, impact assessment, coordinated disclosure, patch regression testing, and deployment confirmation. Access controls must bind models to verified organizations, scoped networks, approved targets, and revocable credentials; human reviewers need enough capacity to reject plausible but wrong exploit paths.
Business implications
Security leaders can now ask for a measurable pipeline: candidates rejected, human review time, private reports, maintainer acceptance, fixes, released patches, severity, and deployment. The pilot suggests the approach can help selected maintainers, but seven researchers spending roughly two days per project is a service model that must be costed before anyone calls it autonomous or scalable.
Where this could be going
Conditional outlook: if similar programs publish severity, review-cost, acceptance, release, and downstream deployment data across more projects, governed AI-assisted remediation could become a standard security workflow. If the candidate-to-fix ratio and expert workload remain this high, it will stay a high-touch capability for selected codebases rather than a general replacement for security engineering.
Risks & uncertainties
- The NATO token figure, institutional-access claims, and ENISA summary are reported by OpenAI; the reviewed sources do not independently break down tasks, model versions, cost, success rate, or avoided incidents.
- Token consumption measures activity, not vulnerability quality, defender advantage, or operational effectiveness.
- CERT Polska independently documents six RouterOS vulnerabilities and the supporting agent workflow, but one successful program does not establish results across software classes or teams.
- CERT Polska says the most labor-intensive work remained context preparation, safe-lab design, research direction, full verification, false-conclusion removal, and impact documentation.
- The RouterOS findings were already under active exploitation by the time of public disclosure, showing that discovery and patching do not eliminate the patch gap.
- The earlier Linux pipeline counts and benchmark results remain vendor-reported and have not been independently reproduced at the same scale.
- External evaluation did not establish universal superiority over general frontier models or success on the hardest fully hardened targets.
- Restricted access, monitoring, and human oversight reduce risk but do not remove dual-use, credential, target-scope, or exploit-leakage risks.
- The MDASH pilot covered six consenting projects, a small and selected sample that does not establish performance across ecosystems or codebase types.
- The 99 reports were private and cannot be independently reconstructed from the reviewed public source.
- Reported fixes do not establish severity, downstream deployment, avoided incidents, or whether unreleased fixes will reach users.
- Seven researchers spending about two days per project shows meaningful human cost and limits claims of autonomous scale.
The AI Signal take
The best number is not 9,791. It is the sequence that follows: 99 reports, 55 fixes, 14 releases. The shrinking funnel is the operating model. AI supplies search volume; researchers absorb false positives; maintainers decide; releases create protection. Anyone selling only the first number is selling noise.
Sources & publication record
GitHub's MDASH account adds first-party operating data from six consenting projects: candidate volume, private reports, fixes, releases, and researcher effort. Private findings cannot be independently reproduced from the public article, so the update treats the funnel as reported evidence and preserves the small-sample, labor, severity, and deployment limits alongside earlier OpenAI, CERT Polska, and Mozilla evidence.
- GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents — OpenAI. Accessed 2026-11-06.
- GPT-6 Cyber System Card — OpenAI Deployment Safety Hub. Accessed 2026-11-06.
- A thousand fewer paths to root — OpenAI. Accessed 2026-11-06.
- Defending the patch gap against automated attackers — Mozilla. Accessed 2026-11-06.
- Security Vulnerabilities fixed in Firefox 159 — Mozilla. Accessed 2026-11-06.
- Strengthening NATO and allied cyber defenses with AI — OpenAI. Accessed 2026-11-17.
- Critical vulnerabilities in MikroTik RouterOS are being actively exploited — CERT Polska. Accessed 2026-11-17.
- Maintainers don't need more AI bug reports. They need fixes. — GitHub. Accessed 2026-11-18.
Published · Last updated
By Dugi Selmanaj. Editorial review recorded in the source data.
View structured content · Read the editorial method