Fortinet discloses nearly 100 flaws after AI-assisted testing | Magica
Fortinet discloses nearly 100 flaws in release with AI-assisted findings
Editorial Team
••📖6 min read
Fortinet's November security disclosures include AI-assisted findings, but some affected software branches still await fixes. Earlier testing put token costs at about $540 per true positive, excluding the broader work of reviewing findings and securing customer deployments.
Fortinet's November disclosures address nearly 100 vulnerabilities, including critical flaws.
AI-assisted findings are part of the release; the advisory total does not measure AI's contribution.
Some software branches still have fixes listed as upcoming, making individual advisories essential.
Fortinet, the cybersecurity vendor that combines networking and security, has disclosed nearly 100 vulnerabilities across its products in a November release that includes findings from AI-assisted testing. The immediate question for customers is which installations need an upgrade, a temporary workaround or a move to another software branch.
The scale extends beyond an experimental testing program. In its 2025 financial results, the Nasdaq-listed company reported $6.80 billion in annual revenue and said it served well over half a million customers with more than 50 enterprise products. Those figures describe Fortinet's overall business, not the number of customers or products affected by these flaws.
A report on the patch release puts the total at nearly 100 vulnerabilities, approximately 17 of them critical and nearly 80 high severity. Several critical flaws could allow arbitrary code execution; others could enable authentication bypass, account takeover or commands running with super-administrator privileges. FortiMail received the largest batch of critical fixes.
Fortinet's product security advisory index shows why severity alone is insufficient for prioritization. It lists CVE-2026-126253, an authentication bypass in the OSPF routing daemon in FortiOS and FortiSASE, as critical and unauthenticated. Another critical flaw, CVE-2026-126229, requires authentication. The index records both internal and external discoveries; an internal discovery label does not by itself establish AI involvement.
JetBrains IDEs 2026.3 bundle Air for running coding agents and reviewing their changes, while the standalone Air app is in maintenance mode. Existing provider subscriptions remain usable, but billing, security controls and execution options vary by integration.
The Center for Internet Security's November 11 advisory rates the risk high for government entities and businesses, and low for home users. It says there are currently no reports of the vulnerabilities being exploited in the wild. That describes available threat intelligence, rather than establishing that every affected installation is safe.
Patches preceded the public advisories
Fortinet's October 21 hardening announcement, updated November 10, scheduled coordinated updates for FortiOS, its operating system, and for FortiManager and FortiAnalyzer on October 22, following releases for FortiMail and FortiAuthenticator. Public advisories were scheduled for November 10. Fortinet says this sequence gives customers time to upgrade before further technical details become public.
The management products occupy consequential positions in a customer's network. FortiManager provides centralized administration of Fortinet devices; FortiAnalyzer supplies centralized analytics and logging, as described in the company's management product data sheet.
The hardening announcement said FortiSASE would update automatically before October 22, while FortiManager Cloud could require a manual update and public-cloud instances would follow cloud partners' release processes. Deployment responsibilities therefore differ even within the same vendor's portfolio.
The broad announcement also needs to be checked against individual advisories. Its table listed FortiManager and FortiAnalyzer 8.0.2 or later as fixed releases. But the November 10 JSON API advisory FG-IR-26-219 lists 8.0.2 among the affected versions for both products and directs users to an upcoming 8.0.3 release or a workaround. The broad table is therefore insufficient to establish protection against every disclosed flaw.
For that vulnerability, the advisory says a crafted HTTP request may let an unauthenticated attacker execute unauthorized code or commands. It credits an internal discovery using Anthropic's Claude Mythos frontier AI model. Crucially, the vulnerability exists only when the JSON API is enabled, which Fortinet says is disabled by default.
The advisory gives these branch-specific instructions for FortiManager and FortiAnalyzer:
Software branch
Affected versions
Listed remedy
8.0
8.0.0 through 8.0.2
Upcoming 8.0.3 or later, or the workaround
7.6
7.6.0 through 7.6.7
Upgrade to 7.6.8 or later
7.4 and 7.2
Not affected by this flaw
No action for this flaw
The workaround disables JSON API access on affected administrator profiles. The advisory also lists affected cloud and BigData variants, with some fixes still upcoming, and says no virtual patch is available. Its narrower conditions and version lists matter more to an exposed installation than the overall vulnerability count.
Fortinet's company-reported June classifications were 40% true positives, 30% false positives and 30% security best practices. These are not a breakdown of November's vulnerabilities. Source: Fortinet frontier AI testing update.
What the earlier AI tests actually measured
The November update says frontier AI testing complements conventional methods. Fortinet claims the models examine code paths, protocol handling and configuration logic more efficiently than traditional fuzzing and manual review, but the release does not quantify how many of the newly disclosed vulnerabilities AI alone found.
Its June 30 testing account provides a more specific, though company-reported, picture. Fortinet said it had used AI in security analysis for years and was combining multiple frontier models with on-premises models.
Testing that emulated attackers without source-code access did not find a significant number of exploitable vulnerabilities, Fortinet said. However, the models were adept at analyzing firmware changes and developing proof-of-concept exploits for known, fixed vulnerabilities. That distinction supports concern about unpatched systems without establishing that the same tests readily uncovered new flaws.
Expanding testing to source code produced more substantial findings. After more than 100 tests across 30 product families, Fortinet classified its then-current results as 30% false positives, 30% security best practices rather than vulnerabilities, and 40% true positives. These were classifications of the June results, not a breakdown of November's vulnerabilities or evidence that 40% of tests each produced a distinct flaw.
Fortinet also reported an average token cost of about $540 per true positive across all severities at that time. That is a narrower measure than the total cost of finding and fixing a vulnerability. The company said skilled staff still had to design prompts and test harnesses, interpret results, address issues and validate fixes. Analysts manually reviewed every finding before accepting a false-positive classification.
The figures establish that Fortinet has disclosed some testing economics. They do not provide a like-for-like comparison with conventional discovery costs or measure the full cost through customer remediation.
The next decision is exposure, version and remedy
There is a relevant history behind the urgency. An October 23, 2024 Canadian government alert said Fortinet had received reports of exploitation of CVE-2024-47575 affecting FortiManager products. That earlier incident does not establish exploitation of the newly disclosed flaws.
The Center for Internet Security recommends applying appropriate updates immediately after testing. Fortinet prioritizes internet-facing devices and advises restricting administrative interfaces to trusted addresses or disabling external administrative access when upgrades cannot happen immediately. Those restrictions reduce exposure but do not replace an upgrade.
Customers must now reconcile their actual versions and enabled interfaces with each applicable advisory, including fixes still marked upcoming. To assess the broader AI program, the missing evidence is a comparable measure of validated discoveries, total staff and computing costs, and time to completed remediation. The November vulnerability count alone cannot answer that question.
Amazon’s shares and convertible notes could be worth $271.8 billion at the top of Anthropic’s proposed IPO range. Existing accounting gains, a 12% ownership cap and Anthropic’s computing commitments shape what that paper value means.
Nvidia chief Jensen Huang argued in May 2025 that US chip restrictions were strengthening Chinese rivals. Filings show a substantial commercial stake, but lost sales and market share alone cannot establish whether the controls failed their security purpose.
Reported demand has Anthropic’s bankers considering a price above its filed $70–$78 range. The combined financing could exceed $100 billion, but existing cash, uneven operating results and long-term computing commitments complicate the valuation test.
The British AI chip designer has secured private and UK government backing before its planned 2027 commercial launch, putting a multibillion-dollar valuation on its promise of faster, cheaper inference.
Zhang Yiming reportedly told staff that ByteDance needs years to challenge leading US AI labs, even as the TikTok parent prepares a personal agent for release within weeks. Its consumer reach and video strengths have yet to translate into frontier language models.
Helsing has unveiled a military intelligence platform promising European control over data, models and operations. Germany's evaluation of roughly 30 European systems makes procurement and demonstrated performance the next tests.
Public safety leaders warn that synthetic voices and background sounds make hoax emergency calls harder to assess. A voluntary survey exposes preparedness gaps, but does not establish a national AI-swatting rate or measure growth over time.
Tamil Nadu told the Madras High Court that accounts containing about 2.22 million rupees had been frozen after its library-procurement order, while a criminal case was still being prepared. The dispute tests safeguards ordered in 2024; AI involvement remains unconfirmed.
Unions withdrew objections to Google’s proposed $10 million purchase of Spirit Airlines data after securing employee-record exclusions and independent privacy review. The agreement follows earlier passenger-data protections; court approval and verification remain ahead.
Baird cut Nvidia to Neutral and its price target to $200 from $500. Its case combines slower forecast AI spending growth with China-linked inventory exposure and uncertainty over founder and CEO Jensen Huang’s tenure.
ByteDance is reportedly accelerating plans to combine its Spell project with Doubao. Existing travel bookings reveal a transaction-fee business, while the phone assistant's history shows how access to other apps can constrain execution.
The Office for Students wants more AI-free assessment and better AI skills teaching, backed by existing quality rules. Its research raises doubts about unsupervised coursework, but universities say assessment reform is already under way.
Finland has recognised Tesla’s provisional Dutch approval after changes to branding and speed controls. Drivers remain responsible, and the national decision does not commit Finland to backing EU-wide authorisation.
Grok Bot can connect to internal MCP servers through Tailscale without exposing them to the public internet. User-based authorization preserves network controls, but administrators must configure how narrowly the assistant can reach into their systems.
Daylight’s Goldman Sachs Alternatives-led Series B backs security agents supported by human engineers. The young company reports rapid growth, but has not disclosed the economics behind its automation promise.
Primordial Labs raised $18 million to widen deployment of Anura, its military command software. Existing contracts and demonstrations give it a foothold, but integration counts do not establish sustained use or lower operating costs.
Northwestern Polytechnical University bought two Nvidia H200 processors for 874,700 yuan and awarded five contracts for rented Nvidia computing, procurement records show. Separate US restrictions on the defense-linked institution remained in place despite January’s conditional opening for H200 exports to China.
The North Dakota Republican says allegations involving a suspected Chinese intelligence operative make Huang vulnerable to coercion. Nvidia’s board is backing its founder while independent directors investigate, with investors facing both security and leadership risks.
Scott Bessent says Washington and Beijing are establishing a channel for serious AI incidents, following their September agreement. Its usefulness depends on unresolved reporting procedures and participation by the companies that detect model failures.
