Use Case · Zero-Day Discovery

Find Zero-Days Pattern Scanners Miss

Pattern-matching scanners can only find vulnerabilities someone has already reported. Safeguard's engine plus Griffin AI traces taint across package boundaries and hypothesises exploit conditions — surfacing candidate zero-days in your dependency graph before they are published anywhere else.

81%
Exploit Hypothesis Accuracy
10k+
Packages Continuously Analysed
<1h
Candidate Review Triage
98%
Adversarial Prompt Resistance

Why Pattern Scanners Never Find Them

Traditional SCA tools are looking for vulnerabilities that already have CVEs. A zero-day, by definition, doesn't.

01

CVE-Based Detection Is Inherently Reactive

Pattern scanners match your dependencies against a database of known vulnerabilities. If a vulnerability hasn't been disclosed, there's no pattern to match. You find out about the zero-day when the attacker uses it on someone else.

02

Cross-Package Taint Flows Are Invisible

Your HTTP handler in app.js reaches a sink in a transitive dependency 7 hops deep. No single file contains the full vulnerability. Pure static analysis in one package misses the flow that crosses package boundaries.

03

Pure-LLM Bug-Hunters Have 60-95% False Positive Rates

LLMs asked to 'find vulnerabilities' flag every risky-looking sink. Without structured context about reachability and exploit conditions, the output is unusable at production scale.

04

Triage Capacity Is Already Saturated

Security teams drown in thousands of low-signal findings. Adding more noise to the queue — even if some of it is signal — doesn't help. What's needed is high-precision candidates, not high-volume alerts.

The Engine-Plus-LLM Pipeline

The Engine Finds The Shape. The Model Hypothesises The Exploit.

Stage 1 — Taint Analysis

The deterministic engine walks the cross-package call graph, propagates taint from every source type to every sink, and surfaces every path that survives existing sanitizers.

Cross-package call graph construction
Version-aware symbol resolution
Sanitizer-aware propagation

Stage 2 — Griffin AI Hypothesis

For each surviving path, Griffin AI receives a structured brief (source, sink, intermediate code, version) and hypothesises the exploit class, trigger input, and CVE mapping — if one exists.

CWE classification
Input hypothesis
Known-CVE deduplication

Stage 3 — Disproof & Verification

A second model pass tries to disprove the hypothesis. Only candidates that survive the adversarial check reach the review queue — keeping precision high and triage hours low.

Adversarial disproof
Ranked evidence
Coordinated disclosure workflow
The Safeguard Model Lineup

Three Weighted Models, One Cybersecurity Brain

Every model is weighted purely on cybersecurity corpora — CVE bodies, exploit write-ups, taint graphs, advisory disclosures, malware behavioural traces — not general internet text. Each is tuned for a different point on the precision/latency curve.

They cooperate: Lion flags candidates at commit time, Eagle widens the surface across a repo, Griffin proves or disproves the hypothesis with deep reasoning.

Griffin

Deep reasoning · the hypothesis engine.

Heavyweight reasoning model. Walks 100-level dependency graphs, hypothesises exploit conditions, runs an adversarial disproof pass, and generates remediation PRs. Best when you need a single high-precision verdict on a candidate path.

~70B params
Multi-step chain-of-thought
~6–10s per finding
98% adversarial prompt resistance

Eagle

Wide-angle triage · the surface scanner.

Fast attack-surface mapper. Sweeps the call graph, clusters taint flows, and ranks the top candidate paths in seconds so Griffin spends its budget on the right ones.

~13B params
Ranking + clustering head
<500 ms per package
Feeds the queue Griffin draws from

Lion

Inline · the commit-time gut check.

Tiny, distillation-trained inline model that runs in the IDE / CLI / pre-commit hook. Catches obvious vulnerable sinks and bad sanitiser usage before code reaches CI.

~1B params
<80 ms p95
Runs locally
No source code leaves the developer machine
How They Hand Off
01Lion
Commit

Inline sink + sanitiser check on the dev's machine.

02Eagle
Repo scan

Cross-package taint paths ranked across the codebase.

03Griffin
Path hypothesis

Deep reasoning posits an exploit chain + CWE class.

04Griffin
Adversarial disproof

Second pass tries to refute; survivors hit the queue.

Anatomy Of A Candidate

What A Zero-Day Actually Looks Like From The Queue

From First Commit To Your Queue

  1. t = 0Lion

    Flags a risky deserialization sink in PR #4129. Inline, <80 ms, on the developer's laptop.

  2. t + 90 sEngine

    Reconstructs the cross-package call graph for the merged commit, version-pinned.

  3. t + 3 minEagle

    Ranks 47 surviving taint paths; 6 reach untrusted HTTP entry points.

  4. t + 8 minGriffin

    Hypothesises an unsafe-deserialization → RCE chain for the top candidate, attaches a synthesized trigger input, maps to CWE-502 with no matching CVE.

  5. t + 9 minDisproof

    Second pass fails to refute under the project's sanitiser config; finding lands in your review queue with full evidence bundle.

What You Actually See

Every finding lands with the full evidence bundle attached. No "trust us" verdicts.

Cross-package taint path

Source → 6 hops → sink, every node version-pinned.

CWE class

Mapped to the closest CWE, even when no CVE exists yet.

Hypothesised trigger input

A synthesized payload that should reach the sink.

Disproof attempt log

Every refutation the adversarial pass tried and failed.

Suggested patch

PR-ready diff with compatibility notes for the pinned versions.

Maintainer contact draft

Coordinated-disclosure email pre-written, opt-in to send.

Blast-radius estimate

Downstream dependents affected if exploited in the wild.

Findings stay private to your tenant. Coordinated disclosure to upstream maintainers is opt-in per finding.

Engine Output

How A Candidate Zero-Day Reaches Your Triage Queue

Taint analysis surfaces a path from HTTP request body into a deserialization sink 6 packages deep. Griffin AI hypothesises an unsafe-deserialization pattern and maps it to a known CWE. The disproof pass attempts to refute the hypothesis with specific mitigation conditions; it cannot. The finding reaches your queue with the full taint path, exploit hypothesis, disproof attempt, and a ranked evidence bundle. If you opt in, Safeguard runs coordinated disclosure with upstream maintainers on confirmed candidates.

6-hop
Cross-Package Taint
~8 min
Pipeline Latency
1-line
Remediation Suggestion

Find What's Hiding In Your Dependencies.

Run the engine against your codebase and see the candidate zero-days your pattern scanner never surfaced.