Zero-Day Discovery

Safeguard runs a continuous zero-day research pipeline across open source packages, container base images, and AI models. Discovered vulnerabilities are validated, scored, and disclosed through coordinated channels — and delivered to customers before they appear in public CVE feeds.

How the Pipeline Works

Zero-day discovery runs through four stages: ingestion, analysis, validation, and publishing.

1. Ingestion

Safeguard continuously ingests:

• Every public commit on watched repositories across npm, PyPI, Maven, NuGet, Crates.io, Go modules, RubyGems, Composer, Conda, Hugging Face, and OCI registries.
• New releases within minutes of publication.
• Maintainer signals — GitHub issues, pull requests, security advisories in draft, and abnormal commit activity.
• Container base image changes — new layers, new binaries, CVE drift.

This ingestion layer is the foundation of the TAOR (Trace, Analyze, Observe, Report) architecture.

2. Analysis

Each new artifact goes through:

• Static analysis — taint tracking, dataflow, symbolic execution, fuzz-driven input generation across multiple analyzers.
• Dynamic analysis — sandboxed execution with syscall / network monitoring to flag behavior that differs from prior releases.
• Differential analysis — diff against the previous version to identify sensitive changes (auth, crypto, deserialization, IPC).
• Dependency graph analysis — 100-level transitive scanning to find whether a change propagates into widely-used packages.
• Griffin AI review — each candidate is reviewed by Griffin AI, which correlates the change against the package's historical patterns, similar vulnerability classes, and known exploit techniques.

3. Validation

A finding becomes a candidate zero-day only if:

1. A reproducer is constructed and runs without manual tweaking.
2. A minimum-qualifying-context (MQC) analysis shows a realistic path to exploitation.
3. At least one human researcher signs off on impact and severity.

Pre-validation filtering rejects roughly 96% of raw candidates. The remaining 4% enter the disclosure queue.

4. Publishing

Validated zero-days flow through two channels simultaneously:

• Customer channel — Safeguard subscribers receive the finding immediately through the ESSCM feed, Portal, and CLI / IDE clients. SBOMs are re-scored. Griffin AI generates remediation PRs where a fix exists.
• Upstream channel — coordinated disclosure to the upstream maintainer with a 90-day default embargo (shorter if active exploitation is observed). Safeguard requests CVE assignment and publishes a detailed advisory after the embargo.

TAOR Architecture

TAOR is the name of Safeguard's zero-day infrastructure.

• Trace — ingest every artifact, every commit, every release.
• Analyze — run the multi-tool analysis matrix.
• Observe — watch runtime behavior and dependency telemetry across customer fleets (with opt-in, anonymized signal).
• Report — publish to customers and upstream coordinators.

Observe is the distinctive layer: telemetry from customer SBOMs and runtime collectors feeds anomaly detection back into the pipeline. Unusual package co-installations, sudden version pinning across the ecosystem, and suspicious registry uploads all surface as leads.

Customer Experience

Zero-days appear in your environment as:

• A new finding in ESSCM with category zero-day and a Safeguard-assigned identifier (SG-YYYY-NNNNN).
• A CVE reference once the embargo ends (typically 14–90 days after initial disclosure to you).
• An automated remediation PR from Griffin AI when a patched version or safe configuration exists.
• A policy alert in the Portal if the affected component violates any policy gate.

Signal-to-Noise Discipline

Not every finding is published. Safeguard suppresses:

• Low-impact bugs that do not cross a trust boundary.
• Findings in unmaintained packages with zero downloads in the last 365 days.
• Issues that require an attacker to already have write access to the application.

Research output targets severity medium and above with a realistic exploit path.

API and Feed

Subscribers can pull the zero-day feed via:

safeguard zero-days list --since 2026-01-01 --severity high

Or from the API:

curl -H "Authorization: Bearer $SG_TOKEN" \
  https://api.safeguard.sh/v1/zero-days?severity=high&since=2026-01-01

The feed includes the SG identifier, affected components, reproducer hash, severity, exploitation status (none, poc, in-the-wild), and upstream coordination status.

Researcher Program

Safeguard's research team also accepts external submissions. Validated findings published under the Safeguard disclosure program earn bounties based on impact and novelty. Contact research@safeguard.sh for submission guidelines.

Related

• Griffin AI — how automated remediation ties into zero-day publishing.
• Vulnerabilities — how discovered issues surface inside ESSCM.
• Policies and Gates — block artifacts with unpatched zero-days at build or deploy time.