Fortinet details critical firewall and FortiManager flaws its AI testing found — patch before attackers diff

On November 10, 2026, Fortinet published the security advisories behind its unusual October "hardening release" — and several of the critical ones say the flaws were found by Fortinet's own engineers using the frontier AI model Claude Mythos. Organizations running FortiGate firewalls (FortiOS), FortiManager or FortiAnalyzer should confirm they are on the fixed versions now, because the technical details are now public and none of the critical advisories offers a virtual patch.

What happened, in plain English

Fortinet makes FortiGate firewalls, which run an operating system called FortiOS, plus two central consoles: FortiManager, which pushes configuration to many firewalls at once, and FortiAnalyzer, which collects their logs. In many companies, FortiManager effectively holds the keys to every firewall.

On October 21, 2026, Fortinet's chief information security officer, Carl Windsor, announced that the company had added "frontier AI models" to its internal product security testing and was shipping a batch of critical and high-severity fixes. He urged customers to upgrade quickly but did not describe the bugs. On November 10 — Fortinet's monthly advisory day — the details came out through the PSIRT (Product Security Incident Response Team) page.

Several of the critical advisories credit Fortinet's product security team and say the flaw was found "with the use of the frontier AI model: Claude Mythos", Anthropic's model. None is listed as exploited in the wild.

Think of it as a building owner who quietly changed the locks three weeks ago and has now published a list of which locks were weak. Tenants who already got new keys are fine. Anyone still on the old locks is now more exposed than before, because the list tells everyone where to look.

Are you affected? What to do now

You should care if your organization — or your managed service provider — runs FortiGate firewalls, FortiManager, FortiAnalyzer (including the Cloud versions) or FortiSASE. If you installed the October 22 releases, much of the work is done, but two details below still need a check.

Critical advisories published November 10 (verified on Fortinet's pages):

• FortiManager FGFM service — CVE-2026-126261, CVSS 9.0 (FG-IR-26-275). FGFM is the protocol firewalls use to talk to FortiManager. An attacker who can reach that port, without logging in, could run code or commands. Fixed in 8.0.1, 7.6.8, 7.4.6 and 7.2.9 or later.
• FortiManager / FortiAnalyzer JSON API — CVE-2026-126170, CVSS 9.8 (FG-IR-26-219). Unauthenticated, but only when the JSON API is enabled on an admin profile (it is off by default). Fixed in 7.6.8 and 8.0.3 — note that 8.0.2, the October fix level, is still listed as affected; the advisory offers a workaround (disabling the JSON API) for the 8.0 branch. The 7.4 and 7.2 branches are not affected.
• FortiAnalyzer OFTP log port — CVE-2026-126262, CVSS 9.3 (FG-IR-26-276), SQL injection that could lead to code execution, and CVE-2026-126260, CVSS 9.0 (FG-IR-26-274), a pre-authentication server-side request forgery. Both also affect FortiManager with FortiAnalyzer features enabled. Fixed in 8.0.2 (8.0.1 for some), 7.6.8, 7.4.12 and 7.2.13.
• FortiManager / FortiAnalyzer dmserver — CVE-2026-126229, CVSS 9.1 (FG-IR-26-244). A logged-in administrator could reach other VDOMs (virtual domains — separate tenants or segments on the same box). Fixed in 8.0.2, 7.6.8 and 7.4.12; for 7.2 and 7.0 the advisory lists no fix and says to "migrate to a fixed release".
• FortiOS OSPF — CVE-2026-126253, CVSS 9.3 (FG-IR-26-267). An attacker on an adjacent network could inject routing announcements without authenticating. Fixed in 8.0.2, 7.6.8, 7.4.13 and 7.2.14; FortiOS 7.0 is affected and must be migrated, a branch the October announcement did not list.

Checklist, in order:

• Inventory every FortiGate, FortiManager and FortiAnalyzer, including ones run by a service provider. Ask your provider in writing which versions they run.
• Compare versions against each advisory, not just against the October table. Watch for FortiManager/FortiAnalyzer 8.0.2 (JSON API flaw) and anything on FortiOS 7.0 or FortiManager/FortiAnalyzer 7.2/7.0.
• Upgrade internet-facing and management systems first. Fortinet's upgrade-path tool is linked from each advisory.
• If you cannot upgrade yet: keep FortiManager and FortiAnalyzer management, FGFM and OFTP ports off the internet and reachable only from your firewalls and admin network; restrict HTTPS/SSH admin access to trusted IP addresses; disable the JSON API on admin profiles that do not need it; enforce MFA for every administrator. Fortinet says these steps reduce risk but do not replace upgrading.
• Review logs for unexpected administrator logins, new admin accounts and unapproved configuration changes, as Fortinet advised in October.

Fortinet has not published indicators of compromise for these flaws; none is reported as exploited.

The expert view

What is genuinely new. AI-assisted bug hunting is not new, and vendors have used fuzzing and static analysis for years. What stands out here is a major network-security vendor naming a specific frontier model in the credits of its own critical advisories. Fortinet's blog also thanks Anthropic, OpenAI, Google and SpaceX for collaboration. It is a concrete, production data point for the "defenders use AI first" argument, rather than a lab claim.

The bug classes are old. The vulnerabilities themselves are familiar: an authorization step on a management protocol that happens before the registration password is checked, injection through a log-ingestion protocol, an output-encoding flaw in a JSON-RPC interface, a routing daemon that does not authenticate properly. What the model contributed, per Fortinet, is coverage — analyzing "code paths, protocol handling, and configuration logic" more efficiently than fuzzing and manual review. Proprietary binary protocols such as FGFM and OFTP are classic places where that pays off, because they get less outside scrutiny than web interfaces.

Why the timing matters. Fortinet deliberately shipped fixes about three weeks before details, and argued in October that AI tools let attackers reverse-engineer patches into exploits in "hours or days rather than weeks", and can chain lower-severity bugs into a severe one. Once advisories name the component and the protocol, comparing old and new firmware gets much faster. Treat November 10 as the start of the exposure clock for unpatched systems, not the end of it.

Exposure reality. Most of these flaws require reaching a management or log port (FGFM, OFTP, JSON API), and the OSPF one needs an adjacent network position. Well-segmented deployments are much less exposed than ones with management interfaces on the internet. But FortiManager is a high-value target precisely because it controls the firewalls.

What is still unknown. Fortinet has not said how many of the November fixes came from AI-assisted testing, and totals differ between outlets: SecurityWeek counted 89 advisories and nearly 100 flaws in the month's release, while other outlets list eight critical CVEs, including FortiOS BGP and high-availability sync issues we could not verify against Fortinet's own pages in time for this article. One critical flaw (CVE-2026-126262) was also reported to Fortinet by a researcher on the XOR team at JPMorgan Chase, so outside researchers are looking at the same code. The FGFM advisory's fixed versions (7.4.6, 7.2.9) are older than the October releases; the advisory does not explain when those fixes first shipped.

Official sources

• Fortinet PSIRT advisories — FG-IR-26-275, FG-IR-26-219, FG-IR-26-276, FG-IR-26-274, FG-IR-26-244, FG-IR-26-267 (all published November 10, 2026)
• Fortinet blog: Fortinet Releases Security Hardening Release for Multiple Products — Carl Windsor, October 21, 2026, updated November 10, 2026
• Coverage: SecurityWeek — Fortinet Patches Nearly 100 Vulnerabilities
• Coverage: Cyber Security News — Critical Fortinet flaws expose FortiManager and FortiAnalyzer
• Coverage: SecurityOnline — Fortinet patches eight critical flaws