Security
Security hub
Section titled “Security hub”Open Security in the workspace sidebar. The hub has Overview, Dependency scans, and Repository audits tabs. Each check displays its own access, loading, and evidence state. A dependency result does not change a repository audit verdict, and an unavailable check does not prevent viewing the other one.
Dependency & CVE Scan
Section titled “Dependency & CVE Scan”Free Beta Plus and Studio or higher plans include Dependency & CVE Scan, powered by OSV-Scanner. Open Security → Dependency scans, upload a lockfile, and select Scan dependencies. Each scan runs in its own isolated E2B sandbox; Celery handles the queue and saves the result. You can reopen the page in the same browser tab to retrieve its latest scan, or download the completed report as JSON.
Supported uploads are package-lock.json, yarn.lock, pnpm-lock.yaml,
poetry.lock, uv.lock, Cargo.lock, and composer.lock. Upload one UTF-8 file
with its original filename, up to 2 MiB. One scan can be active per account, with
up to ten submissions per hour per source IP. Scans require the hosted background
worker, Redis, and E2B. Local-only installations cannot use this online scan.
Free Beta Plus uses the same scan limits as paid plans. When the trial ends and
the account returns to Free BYOLLM, dependency scans require Studio or above.
If E2B is unavailable, scanning fails without a security verdict.
Local, workspace, editable, and direct-file dependencies are not supported. A
lockfile containing them fails with incomplete coverage rather than receiving a
clean verdict. Cargo supports registry dependencies only: workspace/root packages
and Git dependencies are rejected because the scanner cannot retain their source
identity. Classic Yarn lockfiles with escaped strings and Yarn direct-archive
selectors are also rejected. Supported Git dependencies must use a remote source
and identify a full 40-character commit; the report retains that commit.
What the report means
Section titled “What the report means”The scanner compares dependency identities and versions against the public OSV vulnerability database. Findings include CVE identifiers where available, ecosystem advisory identifiers, package versions, severity when reported, and fixed versions listed by the relevant advisory. Several release branches may have different fixes; choose a compatible release and run your tests after an upgrade. Fixed versions are advisory data, not an automatic patch recommendation.
No known vulnerabilities found means OSV returned no matching advisories for the dependencies identified in that file at scan time. It does not prove the application is secure. Failed, expired, empty, and incomplete scans do not receive a clean verdict.
This feature does not scan application source code, secrets, container images, operating system packages, or deployed services. It does not run penetration tests, determine exploitability, or provide scheduled monitoring or automatic remediation. Dependencies absent from the uploaded file are outside its scope. No package installation, build scripts, or call-graph analysis are executed.
Use from an MCP client
Section titled “Use from an MCP client”Connect your agent to the QualityMax MCP endpoint using your account. The
start_dependency_scan tool accepts filename and the lockfile’s UTF-8 content
(not a local path or URL). It returns a queued scan in scan, including its id.
Call get_dependency_scan with that ID as scan_id to retrieve the status.
Poll no more often than every two seconds while queued or running.
Completed responses include the report under scan.result and pagination.
Follow pagination.next_offset using the offset parameter until it is null
to read all findings. limit defaults to 10 and can be at most 20; the byte limit
may produce a smaller page. A finding too large for MCP returns an explicit
error with a dashboard link containing its scan ID. Open that link while signed
in to the account that started the scan, then select Download JSON report.
You can also open /security/dependencies?scan_id=<scan ID> directly, including
from the new workspace design. Other accounts cannot access the report.
Failed scans never imply a clean result.
Both tools require accepted Terms of Service and Free Beta Plus or Studio or above,
and are included in the public, core, and infrastructure catalogs/profiles. Reading a report requires OAuth mcp:read;
starting a scan requires mcp:write. Access is checked again when queued work runs.
MCP allows ten submission attempts per hour per user and shares the dashboard’s
one-active-scan limit. The authenticated account determines report ownership;
agents cannot supply another user ID. Lockfile contents and report contents are
excluded from the server’s MCP call traces. HTTP and stdio MCP analytics also
drop these tool events, including their exception events.
Web and worker error telemetry also removes scan request bodies, report contents,
and traceback locals while retaining error types and stack locations.
Data and access
Section titled “Data and access”The uploaded lockfile is sent to an isolated E2B sandbox. Package names and versions are sent from that sandbox to the public OSV service for vulnerability matching. Do not upload credentials or secrets. Lockfiles are temporarily held in shared job storage for up to 15 minutes and removed when processed. Reports expire 24 hours after their last job update and are available only to the user who submitted the scan, while their plan remains eligible. Temporary scanner files disappear when the sandbox is destroyed after scanning. A five-minute provider lifetime also bounds sandbox retention if the worker crashes. Download a report if you need to keep it.
The downloadable report records the uploaded filename, SHA-256 digest, scanner
version, execution backend (e2b), completion time, package count, and normalized findings. It does not
contain the uploaded lockfile. The page displays up to 100 findings; the JSON
download contains all supported findings. Reports exceeding processing limits
fail explicitly instead of silently truncating coverage.
Repository audits
Section titled “Repository audits”Open Security → Repository audits and choose a connected repository. You can inspect its recent audit history or select Run repository audit. Confirm the request after reviewing the account-credit notice. A completed audit has a 30-minute cooldown; queued and running audits have no security verdict yet. A failed run stays visible alongside earlier evidence.
These AI reviews examine route authorization and tenant isolation. They are separate from dependency scanning and do not prove exploitability or application security. Grades identify high-risk routes, routes needing review, routes with no issue flagged, and public routes. Selecting a grade loads matching routes from the server. Search applies to loaded route paths, files, and methods. The API returns at most 500 findings per selection; the page shows a partial-list notice when that selection contains more routes. History shows up to 30 recent runs for the selected repository.
Open a finding to see its source file, line, audited revision, and review state. Source links are offered only for GitHub receipts with a full recorded commit. Historical seeded baselines are labeled and do not represent a new audit of the current code.
The detail view separates the saved Audit explanation from a human Review decision. Older findings may have no recorded explanation; a new audit is needed to capture one. What this grade means and Review and verify provide general guidance, separate from the model’s observations. Review shared authorization code and database policies, then test allowed and denied access with separate accounts in an authorized test environment. For public routes, confirm that unauthenticated access exposes only intended public data. After a fix, rerun the relevant tests and audit the updated revision.
You can acknowledge a finding or up to 200 selected findings, optionally adding
a reason, and undo an acknowledgment later. Acknowledgment records a review
decision; it does not fix code or remove the risk. The page reloads the stored
finding state after each change. Operator scanner probes remain a separate
staff surface at /security/scanners.