Skip to content

Security and trust

AI release agents
you can hold to account.

AI AEGIS recommends whether a release should ship. Here is how we keep that call honest, keep the agents in bounds, and keep your data where you allow it. Everything on this page describes how the product works today.

Release verdicts

AI can raise a flag. It can never lower one.

The go/no-go call comes from one small, readable rule. Blockers mean NO-GO, conditions mean CONDITIONAL GO, and only a clean list means GO.

  • Rules decide, every time

    Each stage agent runs its rules first and the orchestrator agent rebuilds the decision from the blocker and condition lists. The same evidence always gives the same call.

  • AI review only adds caution

    When AI review is switched on, Claude can add findings, conditions or a hold. It cannot remove any, so a confident model can never turn a NO-GO into a GO. AI review is optional, and the agents work without it.

  • Findings cite their evidence

    A model finding must quote the input it relies on, and the quote is checked in code. A finding whose citation does not check out is marked unverified.

  • Inputs are data, not instructions

    Stories, specs and diffs are passed to the model as untrusted material to evaluate. The review prompts tell it never to follow instructions found inside them.

  • Missing means missing

    Evidence that was not provided or could not be read counts as missing, not as a pass, so gaps show up in the verdict instead of disappearing.

  • A person signs off

    A verdict is a recommendation. A person records the sign-off, and the report can be exported as PDF, Markdown or JSON for the audit trail.

Agents in bounds

Every agent works inside a fence you can read.

The agents analyse and recommend. They never deploy, and they never merge code.

  • An allowed tool list per task

    The stage agents and the orchestrator run inside a harness that only lets each task call the tools listed for it. Anything else is refused before it runs.

  • Each run is traced

    The harness records every agent run, its tool calls and its model token usage as OpenTelemetry spans that can be sent to a collector you choose.

  • Repairs wait for a human

    The CI repair agent drafts a fix, then stops. It opens a draft pull request only after someone with the approver role approves that exact patch, and a changed patch needs a new approval.

Your data

Read what it needs, from the places you allow.

Each connection is switched off until an operator sets it up, and each one only reaches the projects, repositories and hosts on its allowlist.

  • Jira

    Only reads issues and searches, using an API token that can be scoped to read access. Only projects listed in JIRA_ALLOWED_PROJECTS can be imported.

  • GitHub

    Reads pull request checks and files to gather release evidence, using a token that can be read-only. Only repositories in GITHUB_ALLOWED_REPOSITORIES are reachable.

  • OpenAPI documents

    Fetched over HTTPS only, from hosts in OPENAPI_SPEC_ALLOWED_HOSTS, and redirects are never followed.

  • API tests against staging

    Run only against hosts in API_TEST_ALLOWED_HOSTS, without following redirects. Write requests such as POST, PUT and DELETE are skipped unless you turn them on for that run.

  • Tokens stay on the server

    Integration tokens and model keys are set as environment variables on the API server. They are never sent to the browser or stored in the database.

  • Workspaces are separate

    Stories, runs and verdicts belong to a workspace, and every API call is checked against the caller's membership in that workspace.

Sign-in and hosting

Standard sign-in, HTTPS everywhere, secrets out of the code.

Accounts, roles and hosting use well-known providers rather than anything home-grown.

  • Sign-in through Auth0

    People sign in with OpenID Connect through Auth0. The API checks the signature of every access token against the issuer's published keys before it does anything.

  • Roles for every action

    Workspace members are viewers, members, approvers or admins, and the API checks the role an action needs. Approving a repair, for example, needs the approver role.

  • API keys stored as hashes

    Workspace API keys are shown once when they are created. Only a SHA-256 hash is stored, and an admin can revoke a key at any time.

  • Hosting

    The web app runs on Vercel, the API on Render and the database on Neon PostgreSQL, all served over HTTPS. Secrets live in each provider's environment settings, not in the code.

  • No shortcuts in production

    The local development sign-in mode, which skips token checks, refuses to start when the API runs in production.

Testing

Agent behaviour is tested on every change.

Changes to the API and the web app run through continuous integration before they reach the live site.

  • Agent evals in CI

    A known-case regression suite and a DeepEval scorecard check the agents' verdicts on every push and pull request that touches the API.

  • Tests, types and lint

    The API test suite runs against a real PostgreSQL database, and the web app is type-checked, linted and unit tested, including the verdict rules shown in the public demo.

What we do not claim

AI AEGIS does not hold SOC 2, ISO 27001 or any other security certification yet. If your team needs a security questionnaire answered or a closer look at a control before a pilot, ask us and we will walk you through the code that implements it.

Report an issue

Found a security problem? Tell us.

Send us what you found, the page or API endpoint involved, and the steps to reproduce it. We read every report and will follow up with you directly.

While testing, please use only your own account and data, do not access or change anyone else's, and do not disrupt the service for other people.

Our contact details are also published at /.well-known/security.txt.