Security
Understand access and data flow.
filtro.dev hosts the report workflow. You run the Agent that reads authorized repositories and connects to your model provider.
The hosted service and your Agent
The hosted service receives reports, queues investigations and stores results for your team to review. Your Agent connects outward to collect work and return a narrative with evidence references: repository, revision, file path, line range and explanation.
Repository source must not reach filtro.dev's hosted service. The hosted evidence view shows references and explanations, without source-file previews or diffs. Your configured model provider does process source read by the Agent.
You control repository access
Configure a read-only GitHub token for the organization the Agent investigates. The Agent discovers only repositories available to that token, and report text cannot expand the organization boundary. GitHub and model credentials remain in your Agent environment.
The investigator can read repository information but has no shell, application execution or repository write tool. The Agent needs outbound connectivity; it does not require an inbound listener.
Workspace owners and admins can rotate an Agent token or revoke an Agent on the Agent page. Revocation invalidates its access; unfinished work returns to the queue.
Who processes the investigation
Your Agent sends report claims and the source it reads to the model service you configure. Choose a service and data policy appropriate for your repositories.
filtro.dev screens incoming report text and processes the returned narrative and reference metadata into an automated assessment.
The Agent checks final narratives for source disclosures before submission. Those checks do not guarantee that all source text is excluded from model-written prose.
Evidence before rejection
A negative AI conclusion must identify the affected component and applicable version, trace the attack path and cite an inspected blocking control. Matching a route in a client library does not establish how the server enforces access.
A fresh model review challenges the proposed rejection using the original report and inspected source. Missing evidence, disagreement or an incomplete review produces More evidence needed. This additional review uses your configured model account and contributes to its usage.
Your team approves researcher replies and closure. The AI result remains separate from the human decision. These safeguards do not establish a measured accuracy rate or confirm behavior in a running deployment.
Workspace isolation and access
Workspace access uses owner, admin, engineer and viewer roles. Database access restrictions separate customer records, and Agent jobs are tied to the workspace that registered the Agent.
Review decisions and recorded changes include actor and timestamp information. Human review, severity classification and remediation verification remain separate from the automated assessment.
Information held by filtro.dev
| Category | Contents |
|---|---|
| Reports and conversations | Researcher contact details, submitted report text, intake decisions and inbound and outbound messages. |
| Investigations | Narrative, automated assessment, evidence references, claims, uncertainties and operational counts. |
| Review and remediation | Human dispositions, risk classification, deadlines, remediation links and human verification. |
| Workspace activity | Memberships, Agent registrations, billing records and recorded actions with actors and timestamps. |
A researcher may include a proof of concept in a report. It remains researcher-provided text; receiving it does not mean filtro.dev executed it.
Diagnostics, retention and requests
Operational diagnostics exclude report bodies, source, model message bodies and credentials. Investigation activity includes operational counts to help diagnose failures.
To request access, correction, deletion or retention information, email [email protected]. Individual reports cannot be deleted through the app.
Audit history records changes to workspace data. Its append-only behavior is an integrity control, not a retention period or a determination of how a deletion request will be handled.
Agent updates and interrupted investigations
You deploy and update the Agent. Keep it compatible with the workspace's investigation configuration and review connection errors on the Agent page.
Resource limits and model failures can stop an investigation before it reaches a conclusion. A Needs more evidence result records the gap or interruption; check that explanation before retrying. It does not establish that the report was fully analyzed.
Services involved
| Service | Data and purpose |
|---|---|
| Application and database hosting | Stores and processes workspace records for the hosted service. Contact us for current hosting details. |
| Identity service | Account identity, sign-in, organization membership and invitations. |
| Billing service | Subscription, payment and billing information. |
| Email service | Incoming reports and outgoing researcher replies, operated by filtro.dev. |
| Assessment service | Incoming report screening and assessment of investigation narratives and reference metadata. |
| Your configured model service | Report claims and repository source, processed through credentials you control. |
| GitHub | Repository hosting and read access for your Agent. |
| Your optional diagnostics destination | Operational telemetry configured in your environment. |
Protection and contact
The hosted service is addressed over HTTPS. The Agent connects to app.filtro.dev by default. Protect its credentials and local infrastructure. Contact us for current hosted encryption-at-rest details, data locations and contractual provider terms.
PDF exports organize report history for review. They do not certify your organization or establish compliance.
For security questions, email [email protected]. For privacy requests, email [email protected].