Fortinet Patches AI-Found Flaws in FortiOS, FortiManager

Fortinet on Thursday shipped security updates for FortiOS, FortiManager, and FortiAnalyzer that fix critical and high-severity vulnerabilities uncovered by frontier AI models, and told customers to install them before it reveals the flaws.

The fixes land in FortiOS 8.0.2, 7.6.8, 7.4.13 and 7.2.14, and in FortiManager and FortiAnalyzer 8.0.2, 7.6.8, 7.4.12 and 7.2.13. No CVE identifiers or CVSS scores have been published. According to the announcement by Fortinet CISO Carl Windsor, the advisories will appear on the FortiGuard PSIRT portal on November 10, Fortinet’s next Patch Tuesday.

Every FortiOS build from 8.0.0 to 8.0.1, 7.6.0 to 7.6.7, 7.4.0 to 7.4.12 and 7.2.0 to 7.2.13 is affected. FortiManager and FortiAnalyzer are affected across the same 8.0 and 7.6 ranges, and up to 7.4.11 and 7.2.12 on the older branches.

FortiMail and FortiAuthenticator updates went out a day earlier. FortiMail is fixed in 8.0.2, 7.6.7 and 7.4.9, with 7.2 customers told to move to the 7.4 branch, and FortiAuthenticator is fixed in 8.0.4 and 6.6.11. Those FortiMail builds are the same ones Fortinet identified on October 1 as the fix for CVE-2026-104286, a CVSS 9.8 FortiMail flaw that had already been exploited in the wild and added to CISA’s Known Exploited Vulnerabilities (KEV) catalog that day.

FortiSASE is updated automatically. FortiManager Cloud may require a manual update, and public cloud images are pushed through each cloud provider’s release process.

Why is Fortinet holding back the vulnerability details?

Windsor’s argument rests on patch diffing, in which an attacker compares a fixed build with the vulnerable one to pinpoint the changed code, and on how much AI speeds that up. Fortinet says AI tools can now help attackers reverse-engineer a patch and build a working exploit in hours or days instead of weeks, and chain several lower-severity bugs into a single serious exploit. A detailed advisory released alongside the patch would also hand attackers the map.

“The window between disclosure and exploitation continues to shrink,” Windsor said.

That warning comes from Fortinet’s own testing. In a June update, Windsor said the company was working with Anthropic’s Glasswing (Mythos) and OpenAI’s Daybreak (GPT 5.5 Cyber) programmes. Given only the firmware, as an outside attacker would be, the models found few new exploitable bugs, but they proved highly capable of comparing firmware versions and writing proof-of-concept exploits for vulnerabilities that Fortinet had already fixed.

The source code produced very different results. Across more than 100 tests on 30 product families, Fortinet classed 40% of findings as genuine vulnerabilities, 30% as security best-practice improvements and 30% as false positives, with every finding reviewed by a human analyst. Windsor put the average token cost at about $540 per confirmed vulnerability.

This week’s announcement thanks Anthropic, OpenAI, Google and SpaceX for their collaboration. In a post on X, Fortinet said the programme is also shaping a broader quarterly security release cadence. F5 made a similar shift in July, moving to monthly hardened releases after its own frontier AI testing.

Will attackers wait until November 10?

Not everyone accepts the three-week gap. Yordan Ganchev, principal threat intelligence specialist at watchTowr, told Cybernews he expects the flaws to be reverse-engineered and exploited quickly, and argued that announcing major vulnerabilities in advance mainly primes attackers’ LLM patch-diffing tools. He pointed out that the affected products are repeat entries in CISA’s KEV catalog.

Belgium’s Centre for Cybersecurity issued a patch-immediately warning on Thursday, noting that Fortinet’s announcement contains no CVE identifiers, vulnerability classes or exploitation conditions. No exploitation of the new flaws had been reported as of Friday.

Fortinet edge devices have had a hard year. In June, the FortiBleed credential-harvesting campaign exposed logins for tens of thousands of internet-facing FortiGate firewalls, and the FortiMail zero-day followed this month. The release also lands in one of the busiest patch weeks of the year.

Defenders are already working through PHP2Shell, the critical PHP remote code execution flaws whose details leaked before the fix, and which one researcher now claims to have bypassed on PHP 8.5.12. Node2Shell, a Node.js Brotli decoder bug, has a working RCE proof of concept, and eMudhra is quitting the global TLS business after eight of its intermediate CA keys proved trivially crackable. On Thursday, Trend Micro’s ZDI also posted two unpatched CVSS 9.0 Linux KSMBD flaws to a public kernel mailing list.

What do the FortiOS release notes reveal?

The FortiOS 7.4.13 release notes, dated October 22 and covering build 2933, itemise none of the AI-found flaws. The only CVE they list is CVE-2025-54821, along with a resolved bug (1309910) described as a vulnerability that is triggered when the FortiGate connects to a malicious SSH server and sends crafted packets. Fortinet’s own known-issues entry refers to the new builds as the “Frontier AI vul fix” releases.

Two known issues are worth checking before scheduling the upgrade. Bug 1353958 causes browsers to return ERR_EMPTY_RESPONSE for HTTPS sites on non-standard ports such as 8443 when deep inspection is enabled, after upgrading to any of 7.2.14, 7.4.13, 7.6.8 or 8.0.2. Bug 1349258 causes an uninterruptible upgrade from 7.2.13 to 7.4.13 to fail on FortiGate 7000F chassis.

Fortinet wants internet-facing systems upgraded in the October 22–23 maintenance windows, with configurations backed up first and logs watched for unusual administrator logins or unapproved configuration changes before and after patching. Its release notes say FortiAnalyzer and FortiManager should be upgraded before the FortiGates they manage.

Where an upgrade has to wait, Fortinet advises restricting HTTPS and SSH administration to trusted IP addresses or disabling it on external interfaces, keeping FortiManager administration off the internet, enforcing MFA for every administrator and moving management out of band. The company says these steps reduce exposure but do not replace the update.