Vulnerabilities
A CBOM says what cryptography a product uses. Whether any of that cryptography is broken is a separate question. A different party answers it, at a different time, and the answer does not belong in the CBOM. Instead, the CBOM and the answer are linked. A profile has to require the facts that let a responder holding a CBOM match it to a security advisory. A reader must also know what a passing result does not say about safety.
Three statements, written by three different parties
Three kinds of document describe cryptographic weakness. They are often confused with each other, because they travel together and can even share a file. The table sets out who writes each one, what it says, and when it changes.
| Statement | Written by | Says | Is revised when | Identified by |
|---|---|---|---|---|
| CBOM | Whoever produces the product | This product uses these algorithms at these interfaces, and these libraries implement them | The product changes, or someone observes it more accurately | The subject it describes, plus a revision number |
| Advisory | A third party: a maintainer, a body that issues vulnerability identifiers, a research team | This component, in these versions, has this weakness | Knowledge changes. More versions turn out to be affected, or the analysis is corrected | A vulnerability identifier and a component identifier |
| VEX statement | Whoever produces the product, or someone acting on their behalf | This product is affected by that advisory, or is not, and why | Either of the two above changes, or an investigation finishes | A vulnerability identifier and a product identifier, together |
The producer writes two of the three: the CBOM and the VEX statement. This is why people tend to merge those two. They should stay separate, because they change on different schedules.
A CBOM revision records facts about a product at a moment in time, and those facts do not change until the product does. A VEX statement is a judgement about that product, made against knowledge held elsewhere. That judgement can change several times while the product stays the same. If the two are merged, the producer has to republish the whole CBOM every time a vulnerability feed is updated. If the producer does not republish, it is left holding a document that looks authoritative but is out of date in exactly the part a reader relies on most.
Why a verdict does not belong in the document
The Policy Evaluation section already settles the general form of this. A CBOM records facts. Any judgement drawn from those facts is produced separately, by a versioned policy, at the moment someone asks. Post-quantum posture was the case that forced the rule, and the pqcPosture attribute was removed from the baseline because of it.
Vulnerability status is the same kind of thing and goes wrong in the same way. Saying "this interface is not affected" is a conclusion. It draws on facts the producer holds, on knowledge the producer does not own, and on criteria other people revise. Put that conclusion inside the document and it goes stale without anyone noticing. A signed and dated CBOM that says a product is unaffected is worse than one that says nothing at all, because people will believe it.
The methodology therefore applies decision 0002 to vulnerability status. A profile does not require a vulnerability verdict, and does not permit one as an attribute. It requires the facts from which a verdict can be worked out. This is written up as decision 0020.
Two kinds of cryptographic weakness
Joining a CBOM to an advisory works well for one kind of weakness and hardly at all for the other. An implementation defect has an advisory and an identifier to join on. An algorithm or parameter weakness usually has neither. The methodology therefore has to state clearly which kind of weakness is which.
An implementation defect is a bug in code that performs cryptography. Examples are a buffer overrun in a TLS extension parser, a timing side channel in an RSA decryption path, and a padding oracle. A defect like this has a version range, a patch, and an identifier in a feed. It is discovered on a particular date. The day before, the same code was running and the weakness was not yet known.
An algorithm or parameter weakness is not a bug at all. It is a property of a choice: SHA-1 used for signatures, RSA at 1024 bits, a static Diffie-Hellman group, or any classical key exchange once you assume an adversary with a quantum computer. The code may be flawless. There is no patch, because nothing is broken in the way a patch fixes. The remedy is to use something else, and that means changing a configuration, a protocol negotiation, or the product itself.
| Implementation defect | Algorithm or parameter weakness | |
|---|---|---|
| Identified by | A CVE or equivalent, against a range of component versions | Usually nothing. Guidance deprecates an algorithm, but nobody issues an identifier against a product for still using it |
| Communicated by | An advisory, then a VEX statement for each product | Guidance documents and national profiles, addressed to the industry rather than to any one product |
| Remedy | Update the component | Change the algorithm, which may mean a configuration change, a software update or new hardware |
| Found by | Matching the implementing library against a feed | Comparing the disclosed algorithms with current guidance, which is a policy evaluation |
| What the CBOM does | Narrows the search: it says which interfaces use the affected library, and what each of those interfaces is for | It is the only evidence there is |
The last row shows the main difference between the two cases. That difference determines where the working group should spend its effort.
For implementation defects, the machinery already exists and works. There are feeds, identifiers and VEX formats. A CBOM does not replace any of that. It makes the results more precise.
For algorithm weakness there is no machinery at all. No feed announces that a product signs with SHA-1. No patch arrives. A scanner that cannot tell which interface uses which algorithm cannot tell an operator what to change. A CBOM profile exists for this case. It addresses the case through disclosure rules combined with an external policy, not through anything resembling a vulnerability record.
The line between the two is not perfectly clean, and pretending otherwise would mislead. Some protocol-level weaknesses do get identifiers, because there is a specific implementation path to name. The attacks on export-grade ciphersuites and on weak Diffie-Hellman groups were issued CVEs for that reason. What never gets an identifier is the plain case: current, correct, fully patched code using an algorithm that guidance now says to stop using.
The two kinds of weakness are two of the ten cryptographic failure modes. An implementation defect is failure mode 3. An algorithm or parameter weakness is failure mode 1, or failure mode 2 when the algorithm is adequate today but not for the whole protection lifetime of the data. The other failure modes, such as usage, configuration, key management and fallback, are neither a defect with an advisory nor a property of the algorithm. A CBOM joined to advisories does not find them. Lifecycle Data sets out which stage, and which kind of evidence, reveals each one.
What the CBOM contributes that an SBOM does not
Suppose an advisory is published against a cryptographic library. An SBOM answers one question. Is that library in this product, at a version in the affected range? That tells a responder whether to care. It does not tell them what to do.
A CBOM that conforms to the interface disclosure baseline answers the next few questions, because its rules describe interfaces rather than components. Rule I9 names the library implementing each interface. I7 says what kind of interface it is. I6 says who sits at each end. I1 and I2 give the protocol and its version.
Put together, a responder learns which interfaces use the affected library, whether the management interface is one of them, and whether the product initiates or accepts connections at each. Compare the two answers. "OpenSSL 3.0.7 is present somewhere in this product" has to be triaged by a person. "Three interfaces use it, one of them is the administrative login, and the product accepts connections on all three" can be acted on straight away.
Product rule P2 matters more here than anywhere else, for a reason that has nothing to do with cryptography. The management interface is the one a responder most needs to find, and it is the one most often left out of a disclosure. A profile that only checks the interfaces a producer chose to declare can never catch that omission. This is why the baseline insists that a management interface is either declared, or that its absence is stated.
What makes the join possible
None of the above works unless a machine can put the two documents together without a person in the middle. Three things make that possible, and the methodology already requires all three for other reasons.
- An identifier in a scheme the feeds also use. Rule
I9requires the implementing library as a Package URL, and the profile fixes that scheme rather than letting each document choose. Two disclosures about the same library can then be matched. A product name and a marketing version cannot be matched against anything. The Formats section notes that provider identity is the one requirement SPDX carries natively today, throughpackageUrl, for this same reason. - A lifecycle stage, so the responder knows what was described. Every document declares one stage (rule P6). During an incident the responder needs the running deployment. A product record describes what was built, and may name a library the deployment never loads. The operator's
configuredandobservedrecords describe the deployment. Each of their interfaces names the product interface it realises (correspondsTo), and through that join it inherits the implementing library. A profile written for this use case can accept only deployment records, as the Deployment Disclosure profile does. - A link from the statement to a particular revision of the document. A conformance claim already binds a document by its digest. The same mechanism lets a VEX statement record which CBOM revision its analysis was performed against. That is what keeps the analysis checkable after the product is patched and a new CBOM is issued.
The join, end to end
Three rules of the baseline make the join work without a person in the middle. Rule I9 supplies the key. Reference rule R3 says where to look for advisories. Reference rule R2 says where the producer publishes its own statements of whether the product is affected. The CBOM carries no vulnerability data itself.
- Take the library identifiers from I9. The consumer reads the package URL of the library that implements each interface. In the Keycloak product record, the token interfaces name
pkg:maven/org.keycloak/keycloak-services@26.7.3, and the TLS interfaces name the JDK 21 package of the image. - Query the advisory endpoint under R3. R3 (advisories, SHOULD) points to an endpoint and names its access protocol:
web-page,csaf-provider,osv-apiorghsa-api. The Keycloak product record names two. One is the project's GitHub security advisories page (web-page). The other is the OSV API athttps://api.osv.dev/v1/query(osv-api). OSV is a public vulnerability database that a program can query by package URL, and it aggregates the GitHub advisories. A consumer can therefore send the package URLs from step 1 to the OSV API and receive the matching advisories. - Read the vendor's statements under R2. R2 (vulnerability-statements, MUST for a product) points to where the producer states whether the product is affected by an advisory, and why. The pointer is either a fixed VEX or CSAF VEX document, bound by a hash, or a feed or API with its access protocol named. A producer that publishes no such statements must say so with the stated absence
not-published. The Keycloak product record points to an illustrative VEX document,examples/context/keycloak-26.7.3-vex.cdx.json, by BOM-Link and hash. Its statements name the affected components by BOM-Links into the companion SBOM, whose entries carry the package URLs that I9 names. Upstream Keycloak in fact publishes no VEX, so a real record would statenot-published, ascbom-keycloak-product-disclosure-states.cdx.jsondoes. Silence would fail R2, because a consumer cannot tell "none published" from "not addressed".
The join reaches the deployment too. A deployment interface inherits its library from the product interface it realises. In Example Org's observed record, the database connection resolves through the configured record to the JDK 21 package in the product record. The ingress has no product record, so the configured record names its library itself: pkg:generic/openssl@3.5.4. The example cbom-keycloak-deployment-observed-outside.cdx.json shows the limit. An outside assessor cannot see the ingress library, so that record fails I9 on the ingress.
Conformance checks the pointers, not what they point to. A record that names the OSV API passes R3 whether or not OSV holds any advisory for Keycloak. Context Pointers describes the construct in full.
Where the join breaks down
This use case depends almost entirely on one attribute, and that attribute is the only one the baseline lets a producer withhold. Rule I9 is a MUST, but it carries withholdable: true. This means the producer may supply a withheld marker instead of a Package URL. The marker states that the value exists but is not disclosed. A document that does this still conforms.
For procurement, and for supply-chain disclosure generally, that is a defensible trade-off, and the working group made it deliberately. The library and its version are what a supplier is least willing to publish, as Confidentiality explains. If the baseline had demanded them, suppliers would have refused it outright rather than negotiated over it.
For incident response, the same trade-off defeats the purpose. A withheld I9 leaves the responder with an interface that cannot be matched against any feed, and no other disclosure makes up for it. This has two consequences, and both are difficult to accept.
- A profile written for vulnerability response has to remove the permission to withhold
I9. Removing it makes the rule stricter. A profile that extends the baseline is allowed to do this, and the validator already enforces it. The PQC migration profile does exactly this toI9, for a migration reason rather than an incident one. The procurement profile does the same, so that a buyer can match every interface to advisories for the life of the contract. - Once that permission is removed, only a restricted variant of a document can conform to the profile, in the sense of "restricted" that Confidentiality uses. As a result, the consumer under the most time pressure is the one a public document serves least well. Worse, the responder who most needs the data is often the one with no agreement in place at all: a downstream operator reading a vendor's published CBOM in the middle of an incident. The methodology's answer, settled provisionally, is to be honest about this. A vulnerability-response profile is for a restricted audience, and it does not allow I9 to be withheld. A public reader gets its answer through the producer's VEX statements, which baseline rule R2 points to.
Carrying both in one document
CycloneDX allows a BOM to carry a vulnerabilities array, and a VEX document in that format can either stand alone or sit in the same file as the inventory it refers to. The question is not theoretical. A producer can hand over a single file holding both a CBOM and a set of vulnerability assertions, and a consumer can evaluate that file against a profile.
The methodology's position is that the two are joined, not merged. The rule concerns what a conformance verdict (the pass or fail result of checking a document against a profile) means. It does not restrict what a file may contain.
- A profile neither requires nor forbids an embedded vulnerability array. A profile sets the minimum a document must disclose, not a maximum. Forbidding the array would set a maximum.
- A profile's rules do not examine that array, and a conformance verdict says nothing about it. If a document containing vulnerability assertions passes, that means its cryptographic disclosure met the profile. It does not mean the assertions are true, current, complete, or that anyone has assessed them.
- CycloneDX 2.0 also defines a
riskselement, a risk register. Conformance ignores it in the same way. A supplier's product record should not carry one. - A consumer must not read an empty vulnerability array, or a missing one, as a statement that there are no vulnerabilities. A missing vulnerability statement means only that no statement was made. This is the same principle the disclosure states apply at attribute level, where
undeclareddoes not meannone.
The formats also differ here, in the same way that decision 0016 dealt with for conformance claims. CycloneDX can carry both in one file. SPDX 3.0.1 expresses vulnerability assessment through its own security relationships, and a CBOM referenced from an SPDX SBOM sits in a separate document altogether. A statement such as "the CBOM and the VEX are in one artifact" is therefore true in one format and false in the other. A profile must not be written in a way that depends on which format is used.
Algorithm weakness has no identifier
A responder dealing with an implementation defect can cite CVE-XXXX-NNNNN in a ticket. They can suppress it with a justification, track it to closure, and show later what was decided and why. None of that is available for "this interface signs with SHA-1". There is no identifier to cite, so the finding cannot be suppressed with a reason, cannot be deduplicated across reports, and cannot be compared between tools. Two scanners will report the same problem in two ways that do not line up.
Three existing schemes are sometimes pressed into this role. None of them fits.
- CWE classifies the type of weakness. CWE-327 covers a broken or risky algorithm, CWE-326 inadequate encryption strength, CWE-328 a weak hash. These are categories rather than instances. Every SHA-1 signature in an estate is CWE-328, so there is nothing there to track to closure.
- A clause in a guidance document, such as a section of a national technical guideline, identifies the criterion rather than the occurrence. It also changes whenever the guidance is revised, which is exactly the instability the methodology keeps out of documents.
- A CVE raised against the product is a category error, and the feeds are right to resist it. Using a deprecated algorithm in a supported configuration is not a defect in the product's code.
What the methodology offers instead is the policy evaluation: a versioned, dated policy applied to disclosed facts, producing a finding that carries both the policy version and the date it was run. That is reproducible and auditable. It is still not citable the way a CVE is, because the finding is identified by the policy and the facts it was applied to, rather than by a name in a register that other people recognise. Whether that gap ought to be filled, and by whom, is Q59.
What each kind looks like in practice
Both examples use the Keycloak 26.7.3 records that run through this documentation.
An advisory against a library
Suppose an advisory is published against the JDK 21 package that the Keycloak product record names at I9. The SBOM, which the product record points to under R1, confirms that the package is present. The product record goes further. It says that seven TLS interfaces use the package: the HTTPS listener, the management listener, and the connections to the database, the directory, external identity providers, the mail server and the other cluster nodes.
Rule I6 says which end the product takes. Keycloak accepts connections on the two listeners and initiates them to the database, the directory, identity providers and the mail server. So if the defect can only be triggered by a malicious server, the outbound connections are the exposure and the listeners are not.
The deployment records narrow it further. In Example Org's deployment, TLS from clients ends at the ingress, so the product's HTTPS listener is not deployed. The management port runs plain HTTP. The observed record shows the JDK's TLS in use on the database, directory and cluster connections.
A producer that publishes VEX would then state that the product is affected, with a remediation, or not_affected, with a justification: the vulnerable code is not present, or is not in the execute path, or cannot be reached by an adversary, or is mitigated already. That justification vocabulary exists for the same reason the methodology's disclosure markers do. A negative statement is only useful if it carries its reason. Upstream Keycloak publishes no VEX, and its product record says so under R2. None of this analysis is written into the CBOM.
Guidance deprecates an algorithm
The NIST IR 8547 initial public draft deprecates quantum-vulnerable algorithms at 112-bit strength after 2030. No advisory is published against Keycloak for this. No CVE is issued. No feed fires, and no scanner reports anything, because nothing about the product has changed.
The CBOM's authentication and key facts are the only record of which interfaces are affected. The example algorithm-transition policy, applied to Example Org's observed record, lists six such values: RSA-2048 certificates on the database, directory and cluster connections, RS256 tokens on the OIDC interface and the Admin API, and RSA-SHA256 on SAML. The remedy differs by interface, and the records show which it is. Example Org can move the last OIDC client from RS256 to ES256 by configuration. The cluster transport key is generated by the product and its key type is not configurable, so only a product change can replace it.
This is the case the methodology is built for, and it is the case the vulnerability ecosystem does not address.
What a conforming CBOM does not assert
The Conformance section covers this in general terms. Vulnerabilities need separate treatment, because reading a passing result as a statement about vulnerabilities is the most common way a conformance verdict is misread.
A document that conforms to a profile does not assert that the cryptography it describes is safe, current, unaffected by any advisory, or adequate for any purpose. It asserts that the required facts were disclosed in the required form, and nothing more. A profile sets a minimum standard for reporting. Whether the disclosed facts are good news is a judgement. Someone else makes that judgement, separately, against criteria that will probably have changed by the time anyone asks again.
Positions taken
Four questions about vulnerabilities are settled provisionally:
- Q57. A vulnerability-response profile is for a restricted audience, and it does not allow
I9to be withheld. A public reader relies on the producer's VEX statements. - Q58. A conforming document may carry vulnerability assertions in the same file. A profile says nothing about them, and a verdict does not cover them.
- Q59. A weakness found from a CBOM does not get an identifier of its own for now.
- Q60. When a new CBOM revision is published, nothing happens to the statements bound to the previous revision. They stay true about that revision. Whether a consumer still relies on them is a decision for the consumer's policy.