Assess findings with your source code
Whether a vulnerability affects your product depends on two things. Affectra knows the first: which component is in a release, how it gets there, and which vulnerabilities affect its version. Your source code knows the second: whether you call the vulnerable code, with what input, and behind which checks. A VEX statement needs both.
An AI assistant that works in your repository, such as Claude Code or Codex, and is connected to Affectra sees both. It reads the finding in Affectra, reads your code, and proposes a statement with its argument. A person decides.
What goes where
Section titled “What goes where”- Your source code stays with you and your assistant. Affectra never receives it. Only the text the assistant writes into a proposal reaches Affectra.
- Your assistant sends what it reads to its own provider: the code and Affectra’s answers, as it does whenever you use it on your code.
- Affectra checks the graph part of a proposal, not the code part. The edges a proposal cites must exist in your data. What the assistant says about your code is its claim, and the person who reviews it checks it.
Before you start
Section titled “Before you start”- An assistant that works in your repository and is connected to Affectra. To file proposals it needs the member role, or an API key that can propose changes. See Connect an AI assistant.
- The repository checked out at the commit the release was built from, usually its tag. The code on your main branch may differ from what you shipped.
- Someone to review. You can review your assistant’s proposals yourself, unless your organization turned on Four eyes for agent proposals: then someone else does.
Ask for an assessment
Section titled “Ask for an assessment”The quickest way is the Assess a release workflow: pick the assess_release prompt, or install its skill. See Use Affectra’s workflows. Or name the release and say what you expect. For example:
Use Affectra to list the unassessed findings of Gateway 1.4.0. For each one, find out how the component gets into the release, then check this repository at tag
v1.4.0: do we call the vulnerable code, and can an attacker reach it? Propose a VEX statement only where the code shows it clearly. Put the files, functions and commit you looked at in the explanation, and keep the impact statement fit for our customers. List the rest for me with what you found.
The assistant typically:
- calls
unassessed_findingsortriagefor the release, so the known exploited findings come first. Each item is a finding group: one CVE, however many advisories report it; - calls
explain_findingfor the dependency path from your product to the vulnerable component. If advisories disagree about the component’s version, the answer also shows the advisories that exclude it, with the fixed version they name; - calls
similar_assessmentsto reuse a statement you made on another release, and to read why you rejected earlier proposals; - reads the advisory, which often names the vulnerable function or option, and searches your code for it: calls, build flags, configuration, and the input that reaches it;
- calls
propose_vexfor the findings where the code is clear.
What a good proposal contains
Section titled “What a good proposal contains”| Part | Who sees it | What it should say |
|---|---|---|
| Status and justification | Your customers, in the exported VEX | For example Not affected, vulnerable_code_not_in_execute_path. |
| Impact statement | Your customers, in the exported VEX | Why the product is not affected, in terms a customer understands. No file paths, no internal names. |
| Explanation | Your organization, on the proposal | The code argument: the commit, the files and functions it checked, the search it ran, and why the vulnerable code is not reached. |
| Evidence | Your organization, on the proposal | The edges of the dependency path from explain_finding, and for a conflict the excludedBy edge. Affectra checks them. |
For example, for a finding in zlib’s inflateGetHeader:
- Impact statement: The product uses zlib only to compress log files. It never decompresses input, so the vulnerable function is never called.
- Explanation: Commit 4f2a9c1 (tag v1.4.0). zlib is linked by
src/logging/rotate.conly, which callsdeflateInit2anddeflate. A search forinflateinsrc/andthird_party/finds no call. The build links zlib statically (CMakeLists.txt,ZLIB_USE_STATIC_LIBS), so no other program on the device uses this copy.
Review the proposal
Section titled “Review the proposal”Open Proposals. For each proposal:
- Read the explanation, and check it in the repository at the commit it names. Search for the function yourself. This is the part Affectra could not check.
- Read the impact statement as a customer would. It goes out with every export.
- Accept, Accept with edits, or Reject with a note. The assistant reads the note, so it does not propose the same again.
Assess findings with VEX describes the review page.
When the code is affected
Section titled “When the code is affected”If the vulnerable code is reached, the assistant can do two things, each reviewed by a person:
- Propose an Affected statement with an action statement that tells your customers what to do, for example to update to the next release.
- Fix it in your repository. Ask it to update the component in a branch and open a pull request. Your normal review and CI apply.
When you release the fix and upload its SBOM, Affectra matches the new release. The finding is not on it, and you can export the statements of both releases.
Keep it safe
Section titled “Keep it safe”An assistant that reads advisories can also edit files and run commands in your repository. Advisory texts and SBOM fields are written by third parties, and one can try to steer the assistant. Affectra marks such text as data, but an assistant can still be misled.
- Keep your assistant asking before it runs a command, edits a file or calls a propose tool. Do not approve these for the whole session.
- Use a read-only key or a viewer sign-in for questions. Use member access only for a session that should file proposals.
- Let the assistant open pull requests. Do not let it push to your main branch or publish releases.
- Review what it proposes like any other change. A proposal that is wrong but sounds right is the risk to watch for.