I have filed three Fortinet bugs. In each one the product authenticates by reading back a string the caller supplies about themselves. The advisory that comes after inherits the same defect.
The Nefarious Plan of Fortinet
The command is called
get auth. It does not authenticate.
That is the last line of CVE-2024-47575: FortiManager's get auth Command Does Not Authenticate. It is also, with the names changed, the first thing wrong with the other two files on this vendor.
I have filed three Fortinet cases. In each one the product authenticates by reading back a string the caller wrote about themselves. The name on the product is bigger than the check the product performs. That is the whole file.
The check is whatever the caller typed
FortiManager's fgfm protocol runs on TCP port 541. A client opens TLS, presents a certificate, and sends a command called get auth. The server's one enforcement is that the Common Name on the certificate equals the serialno field in the registration body. The researcher who published the exploit mints a certificate with whatever serial number they choose and writes the same string in the body.
Writing the same serial in two places is the entire authentication.
The certificate does not have to be issued against a real serial Fortinet ever manufactured. It has to parse and have a CN. The server does not cross-reference the string against any inventory. The caller agrees with themselves, and the server confirms that they do.
CVE-2026-39808: One curl to Root on the Box That's Supposed to Catch Malware is the same mechanism applied to the appliance that runs malware under supervision and tells your SOC whether the sample is clean. The polling endpoint expects a jid value. FortiSandbox treats that opaque identifier as implicit authorization, on the theory that the submitter is the one holding the job ID.
The design assumption baked into this endpoint is: the job ID is the implicit credential.
The jid is a string the caller provides. There is no validation of its format. There is no session binding the jid to the submitter. There is a shell concatenation at the other end. The result is unauthenticated root on the verdict authority that FortiGate and FortiMail enforce downstream.
CVE-2026-121529: Fortinet's FortiClient CORS-to-RCE Lands Two "PoCs" In Twenty Hours. One Is A Zero-Engagement Boost. The Other Is A README Advertising A Tree That Does Not Exist. is the same shape in a browser. The CWE is 346, origin validation error. The local agent on the endpoint reads an origin header and decides, from the string in that header, whether the caller is a legitimate FortiClient UI or a web page the user happened to open.
The CVE record names nobody.
Three products. Three protocols. Three fields a caller writes about themselves. The sandbox trusts the submitter's job ID. The manager trusts the peer's serial number. The endpoint trusts the origin header. The authentication is the caller confirming their own typing. Three teams, three languages, one shape.
The advisory is an inventory of the same field
After the breach, Fortinet writes the advisory. The advisory inherits the defect.
The FortiManager advisory enumerates three serial numbers observed registering unauthorized devices, two file paths dropped by the intruder tooling, and nine IP addresses. The IOC list is what Fortinet asks the defender to grep their logs for.
FGVMEVWG8YMT3R63is the serial hardcoded in watchTowr's public PoC.
It is also the string in the public Nuclei template. A defender who finds FGVMEVWG8YMT3R63 in their logs is looking at something somebody typed. The advisory cannot tell them whether the typist was the threat actor or a blue team running the test kit, because the protocol does not record that distinction. Fortinet is distributing an IOC list anchored on the same caller-chosen field that failed to authenticate in the first place. The record of the breach uses the mechanism of the breach.
For FortiSandbox, the advisory lands in April 2026. The bug was reported in November 2025. Public PoC repositories appear within days of public disclosure. Every FortiSandbox 4.4.x in the field ran an unauthenticated root shell for five months, and the defenders responsible for those boxes had no CVE number against which to triage. The bug is five months older than the record of the bug. The gap is by design.
For FortiClient, Fortinet publishes the CVE with the CVSS temporal metric E:P. E:P is the vendor-level attestation that proof-of-concept exploit code exists. The CISA ADP block attached to the same record says Exploitation: none, the SSVC value for a vulnerability with neither public PoC nor evidence of in-the-wild use. Fortinet says a PoC exists. CISA says none does. The two public artifacts the vulnerability-intelligence pipeline indexed as PoCs are a Mastodon quote-share that links its own author's WordPress article, and a four-kilobyte GitHub repository whose only tracked file is a README advertising ten uncommitted paths. Both say, in their own bodies, that no exploit exists.
Someone is bullshitting a defender. Fortinet set E:P. Fortinet credited no researcher. The public artifacts that match the claim are empty. The scanner summed them anyway because their names contain CVE-2026-121529, and the pipeline reads names.
This is where Fortinet is not wrong
FortiSandbox has to run as root. The engine detonates malware, mounts filesystem images, and injects agents into guest VMs. An unprivileged process cannot do that work. The root context of the web endpoint is a consequence of the appliance's job, not a sloppiness choice.
Treating an opaque identifier as a bearer token is not a Fortinet invention. Many vendors ship polling APIs that work this way. CWE-346 is a standards-level problem and CORS is not a Fortinet protocol. The system() concatenation in the FortiManager cp shell-out is a pattern every C codebase has produced at least once. None of that is the charge.
The charge is that across three products, each shipping into a different trust position in the network, Fortinet has made the same architectural call. The security boundary is a field the caller supplies. The product reads the field, compares it to something else the caller also supplied, and treats the equality as authentication. FortiSandbox did it on a polling endpoint. FortiManager did it on a registration protocol. FortiClient did it on a browser-reachable local service.
Fortinet is not alone in making that trade on one product. Fortinet is alone, in the three files I have filed, in making it on every product.
The next one comes from the fabric
Open the FortiManager file and read the paragraph about what FortiManager pushes to the fleet. Open the FortiSandbox file and read the paragraph on the Security Fabric trust model. Open the FortiClient file and read the sentence about the PSIRT index responding Could not retrieve this post at this time.
The Fabric is the product. Every Fortinet appliance joins it. Every join is a handshake where one side registers with the other. Every registration is a string the registering side writes about itself. FortiAnalyzer takes logs from FortiGate. FortiMail submits attachments to FortiSandbox. FortiGate reports to FortiManager. Every one of those handshakes has the same structural question: who is this caller, and who do they say they are. In Fortinet's architecture those two questions are the same question.
The next file comes from a protocol between two Fortinet products where the registering side supplies the identity. Look for a command whose name reads like a security verb. get auth. get verdict. get policy. get session. If the server answers the question instead of challenging it, the pattern is there. You do not need an exploit to find it. You need the string the caller types and the field the server reads back.
The next advisory already writes itself
The next Fortinet case file is a fabric-protocol bug. The vulnerable product is one of FortiAnalyzer, FortiMail, or FortiProxy. The protocol is a registration or polling channel between that product and either FortiManager or FortiSandbox. The CWE is 287, 290, or 346: improper authentication, authentication bypass by spoofing, or origin validation error. The CVSS vector will show PR:N and AV:N. The advisory will list an indicator of compromise that is a string the attacker chose. Fortinet will credit either a boutique firm or no one.
To Fortinet: you wrote auth into the name of a command that asks a server to confirm the client's own typing. You wrote tracer-behavior into an endpoint that runs as root and reads the submitter's job ID as a credential. You wrote E:P into a CVE record and credited nobody. You are not shipping bad patches. You are shipping products whose names are bigger than the checks they perform.
Your customers paid for the names. Add a check the caller cannot write.