Skip to content

Security: lomolabs/lomo-codesecurity

SECURITY.md

Security policy

Lomo CodeSecurity is a local security workbench. It runs scanners against source that the operator explicitly authorizes. It is not a sandbox for hostile repositories, and a passing scan is not a guarantee that code is secure.

The current release is a 0.1 source preview. There is no published long-term support or backport policy. Check release status for platform coverage and known issues.

Reporting a vulnerability

Use private vulnerability reporting in this repository, or choose Security → Report a vulnerability. Private reporting is enabled for this source preview.

If a private reporting channel is unavailable, ask a maintainer through an existing private contact for secure reporting instructions. Do not put vulnerability details, working exploit payloads, credentials or private source in a public issue. No response-time or remediation-time guarantee is offered.

A useful private report includes the affected version, platform, reproducible steps, expected boundary, observed effect and a minimal synthetic example. Do not attach runtime state, browser cookies, editor credentials, private repositories or unreviewed logs. A support bundle is diagnostic material, not proof that its contents are safe to publish.

Local service boundary

The default installation exposes only http://127.0.0.1:8080 for the core application. An unprivileged TCP relay forwards to the API on an internal container network. The API, PostgreSQL and runtime scanner have no host-published ports. The relay receives no persistent volume or runtime secret configuration; it necessarily carries authenticated traffic transiently in memory.

The API enforces exact Host and Origin checks, restricts its configured origin to loopback HTTP with an explicit port, and rejects credentials in URL query strings. Browser mutations require a session-bound CSRF value. Browser sessions use HttpOnly, SameSite cookies. Editor capabilities are bound to the exact profile, origin and authority generation, and stored in account-private files. Browser and editor membership can be revoked independently.

Installation secrets live outside the checkout in owner-private runtime state. Do not commit, share or copy that state between installations. The application does not grant Workspace Trust or repair unsafe credential permissions to make a connection succeed. Local account owners, OS administrators and the Docker engine remain part of the trusted host boundary.

Source code and findings

Source access requires narrow, explicit authorization. Selecting repositories in an editor is separate from granting that access. POSIX and Remote WSL source uses read-only mounts; supported native Windows source uses bounded, sealed snapshots. Path containment, source identity and snapshot digests are checked. Unsafe links, unsupported paths, changed identities and incomplete captures are refused rather than treated as clean scans.

Source transfer permits at most 8 MiB per regular file, 5,000 files and 64 MiB of decoded content. Scanner analysis has a separate 1 MiB eligible-file bound. Capturing a larger code file does not make its analysis complete. Immutable Git verification includes tracked history content; workspace exclusions do not remove tracked files from a protected-push check.

Findings and bounded snippets are redacted before local persistence. Only bounded, redacted supported SBOMs become downloadable scanner artifacts. Raw tool output is temporary and deleted after processing. Redaction recognizes known patterns; it cannot guarantee removal of arbitrary secrets. Do not grant source access when retaining even redacted findings or snippets is unacceptable.

Scanner processes use restricted environments and disposable working state. Unsupported source-owned scanner configuration is refused. These controls do not make arbitrary repository content safe to execute or eliminate defects in third-party scanners.

Scanner data and complete evidence

Runtime scanners use an internal network and a read-only validated cache. Scanner preparation runs in a separate updater with network access to reviewed sources and no project-source mount. Requests select reviewed components, not arbitrary download URLs or commands. Cache activation checks content, digests, compatibility and freshness before changing the active selection.

Bundled definitions expire. The preview's maintenance path is refreshed reviewed source, a rebuild of changed inputs and explicit scanner preparation. It does not promise a signed automatic update channel. See scanner maintenance.

Only a completed scan can have a passing deterministic gate. Failed, partial, cancelled, stale or missing evidence cannot authorize a protected push. Finding dispositions and gate-policy changes do not repair incomplete scanner coverage. Local Git hooks can be bypassed; they are not remote enforcement.

Ollama and external requests

Ordinary scans do not invoke AI. Local Ollama is optional. Remote Advisor routes require explicit host/provider authorization, model identity and a fresh single-finding disclosure confirmation. Automatic and bulk remote routing are not allowed. Cancellation cannot recall content already received by a remote provider. Advisor output is untrusted advice and never determines the gate.

Host control and practical limits

The API does not mount the Docker socket or execute arbitrary host commands. Applying source roots, optional components and repository hooks remains an explicit host-owner action. Do not expose the control origin to a LAN or public network as a workaround for connection problems.

Keep runtime state and named volumes when updating. Recovery helpers are not a validated disaster-recovery facility. The source preview has incomplete cross-platform and failure-recovery qualification; consult release status and report failures without bypassing source, credential or completeness checks.

There aren't any published security advisories