Fortinet Patches First, Discloses Later: Five Product Lines Fix Serious Flaws Found With Frontier AI, Anthropic, OpenAI, Google and SpaceX Thanked, and Details Held Until November 10

What makes this batch different

On October 21, Fortinet chief information security officer Carl Windsor wrote on the company blog that a coordinated set of security updates would ship on October 22 for FortiOS, its firewall operating system, FortiManager, its central management platform, and FortiAnalyzer, its log analytics platform. Updates for the FortiMail email gateway and the FortiAuthenticator identity product had already been released that day. The post says these releases fix vulnerabilities identified through Fortinet's frontier AI-assisted security testing program, including critical and high-severity issues, and thanks Anthropic, OpenAI, Google and SpaceX for their collaboration.

The most unusual part is the order of disclosure. Vendors normally publish advisories and CVE numbers alongside patches. This time Fortinet has published only affected and fixed versions, with vulnerability details to follow on its PSIRT page at Patch Tuesday on November 10. Windsor laid out why customers should treat the upgrade as "a priority change rather than a routine maintenance cycle":

• The releases include critical and high-severity fixes.
• AI tools can now help attackers reverse-engineer patches and build working exploits in hours or days, rather than weeks.
• Once updates are public, the changes themselves reveal where the weaknesses were, making unpatched devices easier targets.
• AI can chain seemingly unrelated low-severity vulnerabilities into a more severe, exploitable one.

He added that the same class of AI capability is reaching attackers, particularly as open-weight models improve, and that the window between disclosure and exploitation will keep shrinking.

Affected versions and stopgaps

The fixed versions (check the official blog for the authoritative list):

• FortiOS: 8.0.2, 7.6.8, 7.4.13, 7.2.14 or above.
• FortiManager and FortiAnalyzer: 8.0.2, 7.6.8, 7.4.12, 7.2.13 or above.
• FortiMail: 8.0.2, 7.6.7, 7.4.9 or above; the 7.2 branch must move to 7.4 or later.
• FortiAuthenticator: 8.0.4, 6.6.11 or above.

FortiSASE was updated automatically before October 22, FortiManager Cloud may need a manual update, and public cloud instances are being pushed through each cloud provider's channels. For customers who cannot upgrade right away, Fortinet recommends restricting HTTPS and SSH admin interfaces to trusted IP addresses or disabling them on external interfaces, keeping FortiManager unreachable from the internet, enforcing multi-factor authentication for all admin accounts, and watching for unusual admin logins and configuration changes before and after patching. It stresses that these steps reduce risk but do not replace upgrading. The post does not say whether any of these flaws have been exploited.

The real numbers behind AI bug hunting

The patches come out of a testing program Fortinet described publicly in June. Windsor wrote then that the company was working with several vendors' frontier cyber models, naming Anthropic's Glasswing (Mythos) and OpenAI's Daybreak (GPT 5.5 Cyber) projects, alongside its own on-premises models.

Testing had two stages. The first emulated an attacker without source code and analyzed firmware only. It turned up few new exploitable vulnerabilities, but the models produced the highest-quality threat models Fortinet had seen and were extremely good at diffing versions and writing proof-of-concept exploits for already-fixed bugs, which is exactly why Fortinet is now patching first and disclosing later. The second stage opened up the source code. As of June, more than 100 tests across 30 product families had produced findings that were 40% true vulnerabilities, 30% security best-practice improvements that were not vulnerabilities, and 30% false positives; security analysts reviewed every finding before accepting a false-positive classification. By Windsor's figures, the average token cost per true positive, across all severities, was about $540. He also argued that frontier models act as "force multipliers" for security teams and are unlikely to replace security professionals any time soon.

For companies running Fortinet equipment, this means patch cadence has to change. For the industry, it is one of the first cases of a major security vendor publicly crediting AI labs and changing its disclosure process because of the risk that AI can reverse-engineer patches. Specific vulnerability types and CVE numbers will come with the November 10 advisories.