NEW RESEARCH: Your Sandbox Is Made of Glass

Read

Trinitite

PricingResearchBlogPodcasts

Skill Vault

Know every skill your AI uses.

Every skill, plugin, and tool your agents use is registered, scanned, and tracked. Unapproved skills are quarantined before they touch production.

Skill Vault (also written skills vault or skillvault) is a signed SKILL.md registry and agentic SBOM for AI agents.

The lifecycle

From download to signed — or quarantined.

A SKILL.md file can come from anywhere — GitHub, Cursor, a community hub. Today most teams load them with no check at all. That is how a single bad skill teaches every agent to leak secrets.

Skill Vault gives each skill a fingerprint, scans it, and only signs it if it is clean. Anything suspicious is held for a human. Your agents load signed skills only.

Every SKILL.md, scanned & signed

agentic SBOM

GitHubCursorCommunityuntrusted sourcesIngestSHA-256Scanbad patternsSignedagents can loadQuarantinedheld for review

How easy it is

Add a skill. Add a rule.

ingest-skill

json

POST /v1/skills
{
  "name": "Kubernetes Deployer",
  "source_url": "github.com/example/k8s-deploy",
  "source_hub": "github",
  "content": "---\nname: Kubernetes Deployer\ntools: [kubectl, helm]\n---\n# Deploys apps to Kubernetes..."
}

Ingest a skill from any source. Trinitite hashes it and queues it for a scan.

custom-rule

json

POST /v1/skills/scan-rules
{
  "name": "Block crypto mining",
  "rule_type": "pattern_match",
  "pattern": "\\b(xmrig|minergate|coinhive)\\b",
  "severity": "critical"
}

Add your own scan rules. A match sends the skill straight to quarantine.

How It Works

Three Steps. Every Skill.

01

Skill submitted

A new skill, plugin, or tool is added to your agent ecosystem — by a developer, a vendor, or an agent itself.

02

Auto-scanned for risks

The skill is automatically scanned for malicious behavior, data leaks, and policy violations.

03

Approved or quarantined

Safe skills get approved. Risky skills are quarantined until a human reviews them. Nothing slips through.

Live Demo

Scan a Skill

Skill Vault — Registry Scanner

3

2

web-search

v2.1.0

12

APPROVED

code-executor

v1.4.2

87

QUARANTINED

email-sender

v3.0.1

25

APPROVED

file-manager

v1.0.0

45

SCANNING

database-query

v2.3.0

18

APPROVED

shell-access

v0.9.4

92

QUARANTINED

Capabilities

Complete Skill Governance

Central Skill Registry

One place to see every skill your agents can use. Who made it, what it does, when it was last scanned, and who approved it.

Automated Security Scanning

New skills are scanned before they go live. We check for data exfiltration, unauthorized network calls, and hidden behaviors.

Approval Workflows

Set up approval chains. A skill goes from submitted to reviewed to approved — or quarantined. No unapproved skill touches production.

Risk Scoring

Every skill gets a risk score based on what it can access, what it does, and how it behaves. High-risk skills are flagged instantly.

Shadow Skill Detection

Find skills your agents are using that nobody approved. If an agent downloads a plugin you don't know about, you'll see it here.

Cryptographic Tracking

Every skill version is hashed and tracked. If a skill changes — even one line — you know. Tamper-proof records for every version.

Every

Skill Tracked

Auto

Scanned

Risk

Scored

Ready

Quarantine

Use Cases

What Customers Protect

Supply Chain Security

Know exactly what code your agents are running. Every skill is tracked from source to deployment with a full chain of custody.

Vendor Risk Management

Third-party skills are scanned and scored before your agents can use them. Hold vendors to the same standards as your own team.

Compliance Audits

Show auditors a complete registry of every skill, its risk score, its approval status, and its change history. All in one place.

Skills Dashboard

One View of Your Entire Skills Library

A single view of your entire AI skills library: what's approved, what's quarantined, what's pending review, and what's been revoked. Track ingestion velocity and scan results over time.

Skills Vault Overview

0Total Skills

Approved

0

Quarantined

0

Pending Review

0

Revoked

0

Ingest Timeline

Week 1

12 ingested

Week 2

18 ingested

Week 3

9 ingested

Week 4

24 ingested

Recent Scan Results

data-transform-v2

clean

email-sender-agent

suspicious

chart-generator

clean

db-admin-helper

malicious

pdf-parser

clean

Quarantine Flow

Suspicious Skills Are Stopped Automatically

When a skill scan finds suspicious patterns — external HTTP calls, root-level permissions, credential access — the skill is automatically quarantined. Human review decides: approve with monitoring, or revoke permanently.

Skill Quarantine Workflow

Submitted

Scanning

Findings

Quarantined

Human Review

Decision

db-admin-helper

Submitted

Source: internal-registry · Version: 1.3.0 · Author: team-data

FAQ

Skill Vault, answered

What is Skill Vault?

Skill Vault (also written skills vault or skillvault) is a registry for the SKILL.md files your AI agents load. When a skill comes in — from GitHub, Cursor, or a community hub — Trinitite hashes it, scans it for dangerous patterns, and either signs it as safe or holds it in quarantine for review. Your agents only load skills that have been signed, and you get a bill of materials (an agentic SBOM) for every skill in use.

What does the scan look for?

The scanner checks a SKILL.md for known-dangerous behavior: piping downloads straight into a shell, base64-obfuscated commands, reverse shells, reading credentials or secrets, and exfiltrating environment variables. You can add your own rules too — for example, block any skill that references crypto-mining tools. A clean skill is auto-approved; a suspicious or malicious one is quarantined.

Does it work with skills from GitHub and other hubs?

Yes. You can ingest a skill from any source by giving Trinitite its source URL and content. Every version is hashed, so if a skill changes even one line after approval, you know. It works with any tool a skill declares it needs, and every state change is written to a tamper-evident audit trail.

Your agents use skills. Make sure you trust every one.

Thirty minutes against your real agent skill catalog. Leave with a cryptographic SBOM and a scoped approval workflow.