Built for AI assistants
Affectra is designed to work with AI assistants from the ground up. Triage, assessment and fleet questions involve a lot of reading: SBOMs, dependency paths, advisories, earlier statements. An assistant can do that reading for you. What it must not do is decide on its own what you tell your customers and authorities. Affectra is built around that split: the assistant reads and proposes, a person decides.
You bring the assistant: Claude, Codex, or any client that speaks the Model Context Protocol. Connect an AI assistant shows how.
Your data as a graph
Section titled “Your data as a graph”Affectra keeps your data as one graph with a published model: products, variants, releases, their SBOMs and components, findings, VEX statements, units and what they report, and the public records behind them, such as CVEs, advisories, KEV entries, EPSS scores and weaknesses. Every node has a stable ID, and every edge says how two nodes are related.
An assistant asks the graph directly instead of piecing pages together. Which of our units run a release affected by CVE-2024-3094, and has the maker assessed it? is one tool call, and the answer names each release, unit, component and statement it used. The assistant can quote those IDs back to you, so you can check its answer in the app.
Tools made from the same model
Section titled “Tools made from the same model”The tools an assistant gets over MCP come from the same model as the graph, so they follow it as it grows. They answer the questions of the daily work:
- What to look at first. Open findings of a product or release, in a fixed order with the inputs of the order: known exploited, KEV due date, affected units, EPSS, CVSS. No opaque risk score.
- Why a finding is there. The dependency paths from the SBOM’s root to the vulnerable component.
- Where else it matters. The releases and units a vulnerability reaches, and statements you already made for it on other releases.
Your assistant can combine these answers with what it knows from elsewhere. An assistant that works in your repository reads the dependency path from Affectra and your code side by side, and tells you whether you call the vulnerable function. Assess findings with your source code shows how.
Affectra tests the tool descriptions with an assistant against a fixed set of questions with known answers. A change that makes the answers worse is not shipped. MCP server lists every tool.
Proposals, backed by evidence
Section titled “Proposals, backed by evidence”An assistant never records a VEX statement, a package mapping or a link between units and a release. It files a proposal for a person in your organization:
- Evidence. A proposal names the edges of the graph it rests on. Affectra checks that each one exists in your data when the proposal is filed. A claim that cites data you don’t have is refused.
- Carry-overs. When the same vulnerable component is reached the same way in another release, the assistant can propose to carry an earlier statement over. Affectra accepts that basis only when the dependency paths match exactly. You review such proposals as one batch.
- Feedback. When you reject a proposal, you say why. The assistant reads your note the next time, so it does not propose the same again.
- Freshness. A proposal can’t be accepted once the data it rests on changed after it was filed, or once it is too old (90 days by default).
- Four eyes. Your organization can require a second person: then nobody decides what their own assistant proposes, whether it signed in as them or uses a key they created.
An accepted statement is an ordinary VEX statement, with a link to the proposal it came from. Assess findings with VEX shows the review.
Access made for assistants
Section titled “Access made for assistants”- Its own credentials. An assistant signs in as you, with your role, or uses an API key marked as used by an AI agent. Keys start as Read only. Only a key that can propose changes, or a person with the member role, gets the propose tools.
- Its own limits. Tool calls count against their own budget per key or person, so an assistant in a loop doesn’t slow down your uploads and collectors.
- Visible use. The API keys page shows each key’s tool calls over the last 7 days and when it was last used.
- Your organization only. An assistant sees your organization’s data and nothing else. Another organization’s ID answers as if it did not exist. Data isolation explains how.
Text from third parties is data
Section titled “Text from third parties is data”Component names and advisory texts are written by SBOM authors and advisory publishers. Someone can put a sentence there that tries to steer an assistant. Affectra returns such text in named fields, never mixed into its own summaries, and the tools tell the assistant that it is data, never an instruction. Because an assistant can still be misled, it only proposes, and a person decides.
Your assistant, your model
Section titled “Your assistant, your model”Affectra itself calls no AI model and sends no data of yours to one. Your assistant reads what its credentials allow and sends it to its own provider, under your agreement with that provider. You choose the assistant, the model and where your data goes.