Skip to content

Universal Source-First Security Research

SecOpsAI now uses one adapter-driven research pipeline for package, artifact, repository, extension, advisory, and campaign investigations. The crates.io workflow remains supported as a compatibility path, but it is no longer a separate product surface.

Start an investigation

Use the canonical command for an exact package or artifact:

secopsai research investigate \
  --ecosystem npm \
  --research-type package_compromise \
  --package suspicious-package \
  --version 1.2.3 \
  --source-reference https://vendor.example/advisory \
  --json

The same command accepts pypi, crates, packagist, go, maven, nuget, rubygems, open-vsx, huggingface, github, container, and chrome-web-store. Use --artifact for an approved local archive when the source is metadata-only or a registry cannot safely provide a download.

research rust-package remains a compatibility alias for existing operators. research package is also a universal compatibility entry point; use --legacy-report only when an older source-backed report file is required. New dashboard actions use research investigate so all ecosystems share the same validation, evidence, model-routing, and publication gates.

Focused aliases use the same implementation: research advisory for vulnerability records, research extension for Open VSX/Chrome extension artifacts, research github-incident for repository/token incidents, and research source for a general source-backed lead. They do not create a second scanner or a separate evidence model.

For a bounded local exposure check, add --search-root or --lockfile. The result reports matching manifest/lockfile lines without importing or installing the dependency. A local path is accepted only by the CLI; browser routes use typed server-side inputs and never execute a path supplied by the page.

Shared safety boundary

Adapters use official allowlisted HTTPS metadata and artifact hosts. Archives are hash-identified and inspected in quarantine with path, size, entry-count, redirect, and expanded-content limits. SecOpsAI never installs a dependency, runs a lifecycle/build/import hook, activates an extension, starts a binary, or submits a sample to a sandbox automatically. Model jobs receive minimized rule-hit context, never raw artifact content.

Sources are recorded separately from attacker IOCs. Registry URLs, vendor reports, GitHub repository links, and documentation domains are references; only validated infrastructure, hashes, and behavior-derived indicators become IOC candidates.

Evidence and publication

An investigation can create or update a Research Case, attach registry and artifact evidence, build an evidence matrix, queue the explicitly selected OpenCodex model, and create a review-only Blog Ops draft. Model output is a bounded review proposal, not proof. Dynamic sandbox submission, disclosure, publication, and deployment remain approval-gated and separate.

Each result reports the normalized research type, ecosystem, package/artifact, version or revision, source references, static findings, validated IOCs, local usage status, missing evidence, route, allowed/blocked actions, confidence, and the explicit execution_performed: false safety state.

Dashboard workflow

Open Administration → Automation → Research pipeline and use Source-First Artifact Research. Select the research type and ecosystem, enter the package or artifact and version, optionally add a comparison subject and source, then:

  1. Preview metadata.
  2. Run safe research.
  3. Build the evidence matrix.
  4. Queue the selected model.
  5. Open the Research Case and review limitations.
  6. Create a review-only draft when publication blockers are cleared.

The Artifact Fleet controls in the same workspace operate the staged metadata → static scan → minimized triage → analyst queue funnel. There are no parallel Rust/npm/PyPI panels.

Capability notes

Registry adapters are active for npm, PyPI, crates.io, Packagist, Go, Maven, NuGet, RubyGems, Open VSX, GitHub repository metadata, and Hugging Face metadata. GitHub and Hugging Face return an explicit metadata-only result unless the operator supplies a reviewed archive. Containers and Chrome extensions can be analyzed from an approved local artifact. Unsupported or unconfigured sources return a visible metadata_only/not_configured result instead of silently falling back to another ecosystem.

Research writing

Use the case evidence and matrix to write reports that distinguish:

  • SecOpsAI deterministic static observations.
  • Registry and repository provenance.
  • External-source corroboration.
  • Imported sandbox observations, if an analyst approved them.
  • Analyst inference and unknowns.

Do not state that a payload executed, credentials were stolen, or a campaign is attributed unless that fact is supported by the corresponding evidence type.