ConsultingWorkbench-backed AI security engagements — map, attack, defend, and prove your AI systems.
Scope a Review
157Scenario files
8First-class tools
16Lab modules
15+ATLAS techniques

Lab modules

The labs are organized as a product, not a stash of experiments.

Module

Prompt Injection and RAG Security Lab

Direct and indirect prompt injection, retrieval poisoning, context leakage, source attribution, access control, and evidence capture.

18 prompt-injection scenarios and 12 RAG-poisoning scenarios in the upstream catalog.

Public-safe lab lineOpen
Module

Agent Security Control Lab

Tool-call authorization, sandboxing, approval gates, telemetry, rollback, audit trails, and blast-radius management.

15 excessive-agency scenarios, 8 agent-abuse scenarios, and delegated-authority coverage.

Public-safe lab lineOpen
Module

Governance Evidence Lab

Map NIST AI RMF, ISO 42001, OWASP LLM Top 10, MITRE ATLAS, and internal AI policy to engineering artifacts.

Governance-bypass, supply-chain, and evidence-capture signals are represented across the upstream package.

Public-safe lab lineOpen
Module

AI Control Crosswalk

Unified framework navigation across OWASP LLM Top 10, NIST AI RMF, MITRE ATLAS, and ISO 42001 — with directional cross-framework mappings, evidence prompts, and scorecard bridges.

Cross-framework graph with directional mappings, confidence scores, and public-safe evidence links across 4 frameworks.

Public-safe lab lineOpen
Module

AI Trust Scanner

Public website trust intelligence for AI claims, legal clarity, security trust, governance evidence, and cross-page consistency.

ATG public scorecard contract, five public dimensions, required caveats, and a Savvy runtime boundary.

Public-safe lab lineOpen
Module

Model Supply Chain Security Lab

Model provenance, registry controls, artifact integrity, model signing, weights protection, and deployment gates.

10 supply-chain scenarios plus model-integrity and sandbox-escape coverage.

Public-safe lab lineOpen
Module

Scorecard Launch Surface

Launch into the public AI Security Scorecard and the organizational AIPSA scorecard when you need the maturity model.

Framework-backed remediation guidance and public-safe scorecard navigation.

Public-safe lab lineOpen
Module

Hiring Calibration Lab

Decompose AI security roles, rewrite job descriptions, build interview loops, and design practical skills validation.

Backed by the report’s role-architecture work and the public labs roadmap.

Public-safe lab lineOpen
Module

Job Navigator

Career onboarding, role targeting, and live AI Security Engineering hiring-signal navigation in one Labs surface.

Thousands of job descriptions indexed and mapped to AI security archetypes, with local profile calibration for candidate navigation.

Public-safe lab lineOpen
Module

NIST NICE Cyber Workforce Role Explorer

Standardized workforce framework explorer for AI Security Engineering roles, KSAs, and task-based benchmarks.

Mapped to SP 800-181 (NIST NICE) with custom AI security extensions.

Public-safe lab lineOpen
Module

LLM Attack Range

Scenario execution, generation/media abuse tracking, and control-evidence readiness in one dedicated product surface.

The flagship product boundary with curated snapshots and public JSON endpoints.

Public-safe lab lineOpen
Module

Prompt Security Reviewer

Deterministic prompt policy review across 15 rules: injection resistance, identity anchoring, output handling, secret detection, and RAG context isolation. Includes KB/corpus scanner.

15 deterministic rules, 9-detector secret scanner, corpus scan over arbitrary document sets — no LLM calls.

Public-safe lab lineOpen
Module

RAG Security Analyzer

Paste a RAG pipeline JSON config and get instant findings across authorization gaps, tenant isolation failures, over-retrieval, document provenance, and sensitive context exposure.

10 rules covering OWASP LLM03 and LLM06 — authorization, tenant isolation, provenance, retrieval scope, and poisoning risks.

Public-safe lab lineOpen
Module

Agent Tool Permission Analyzer

Analyze agent tool configurations for overly broad scopes, missing human-approval gates, unsafe side effects, ambiguous tool descriptions, and privilege escalation paths.

Rules mapped to OWASP LLM06 (Excessive Agency) — write_broad, admin scopes, external transfers, code execution, and identity anchoring gaps.

Public-safe lab lineOpen
Module

Output Handling Safety Tester

Paste model output, select the sink type (HTML, Markdown, JSON, tool call, email, DB, code), and get deterministic safety analysis across injection, leakage, and side-effect risks.

8 sink types, 15 output rules mapped to OWASP LLM05 — script tags, event handlers, unsafe links, command fields, and hidden context leakage.

Public-safe lab lineOpen
Module

Prompt Injection Test Harness

A structured library of 12 attack probes across 10 categories. Record blocked/detected/degraded/passed outcomes per probe and export a full evidence session as JSON or Markdown.

12 probes: direct/indirect injection, system prompt exfiltration, role confusion, policy bypass, tool misuse, data exfiltration, multilingual and encoding bypasses, Markdown injection, multi-turn setup.

Public-safe lab lineOpen

Public trust product

AI Trust Scanner

Public website trust intelligence for AI companies, security buyers, and governance teams.

Dimensions

6

Public surface, AI language, legal clarity, security trust, and consistency.

Artifacts

8

Trust center, legal, contracts, methodology, report, and contact signals.

Caveat

1

Required public-signal caveat wherever score outputs appear.

Use case

Customer trust readiness

Evaluate whether public trust artifacts support enterprise buyer review.

Use case

AI claim-readiness

Review public AI security and governance claims before they become sales copy.

Use case

Vendor trust triage

Create directional public-signal snapshots before deeper private assessment.

ATG scorecard shell

public_shell

The web repo publishes the Labs shell, route, copy, contract, and public-safe sample output.

Based on public website signals and observed artifacts, not proof of any organization's internal security maturity.

Runtime boundary

Rust and WASM scanner work stays in ../savvy-cli using the existing pure-engine plus thin-WASM-wrapper pattern.

Flagship product

LLM Attack Range

Scenario execution, attack-pack coverage, and control-evidence readiness in one dedicated product surface.

Scenario files
157

Upstream scenario catalog curated into the public product surface.

First-class tools
7

promptfoo, garak, PyRIT, AgentDojo, Giskard, Inspect AI, and NeMo Guardrails.

Screenshot-backed scenarios
19

Scenarios with explicit screenshot capture in the upstream package.

ATLAS mappings
15+

Distinct MITRE ATLAS techniques represented in the scenario catalog.

multimodal_exfiltration

097-image-steganography-exfil

Severity critical · deepfakes_synthetic_media

Success 55% · Controls 44% · Evidence 81%

prompt_injection

001-prompt-injection-basic

Severity high · prompt_and_generation_security

Success 42% · Controls 68% · Evidence 88%

synthetic_media_abuse

154-deepfake-script-generation

Severity high · deepfakes_synthetic_media

Success 39% · Controls 52% · Evidence 86%

multimodal_jailbreak

100-video-frame-injection

Severity high · prompt_and_generation_security

Success 34% · Controls 61% · Evidence 84%

Boundary

AI Security Range

curated_snapshot

Curated from the upstream local-first lab package and imported into this repo as a public product shell.

ATLAS mappings

15+

OWASP mappings

11

Run window

2026-01-01 to 2026-04-30

Directional lab signal built from synthetic scenario runs, attack-pack coverage, and control-evidence rollups.

Public assets

The labs page exposes the curated data bundle directly.

Ready

Catalog snapshot

Versioned product catalog for the public labs shell.

Size: 1 JSON file

Ready

Scenario snapshot

Curated scenario rows and control-evidence rollups.

Size: 157 scenario rows

Ready

Overview snapshot

Status, scope, caveat, run window, and data endpoints.

Size: versioned overview

Ready

Metric bundle

Public-safe scorecards and range metrics.

Size: 4 metric cards

Ready

Generation / media rollup

Monthly synthetic-media and prompt-abuse tracking.

Size: 8 monthly rows

Ready

ATG sample scorecard

Public-safe AI Trust and Governance scorecard sample with required caveat.

Size: 1 JSON file

Included boundaries

  • • Scenario snapshot, overview snapshot, metric bundle, and generation rollup.
  • • Trust Scanner scorecard sample with required public-signal caveat.
  • • Chart exports for generation attacks, output safety, control effectiveness, and evidence capture.
  • • Public README describing the product shell and what stays upstream.

Roadmap

The lab product expands into repeatable validation tracks.

Lab track

Prompt Injection and RAG Security Lab

Direct and indirect prompt injection, retrieval poisoning, context leakage, source attribution, access control, and evidence capture.

Direct prompt injectionIndirect prompt injectionRetrieval poisoningContext leakageEvidence capture

Lab track

Agent Security Control Lab

Tool-call authorization, sandboxing, approval gates, telemetry, rollback, audit trails, and blast-radius management.

Tool authorizationApproval flowsTelemetryRollbackAudit trails

Lab track

Governance Evidence Lab

Map AI RMF, ISO 42001, OWASP LLM, MITRE ATLAS, and internal policy to engineering artifacts.

Control mappingEvidence registryRisk registerAudit trailReport generator

Lab track

Model Supply Chain Security Lab

Model provenance, registry controls, artifact integrity, model signing, weights protection, and deployment gates.

ProvenanceArtifact integritySigningDeployment gatesModel cards

Lab track

Hiring Calibration Lab

Role decomposition, JD rewrite, interview scorecards, lab-based screens, and skill matrix design.

Role architectureJD rewriteInterview loopsSkills validationCandidate rubric

Provenance

What this repo publishes and what it does not.

The executable lab runtime remains upstream; this repo publishes the product shell, curated snapshots, and public-safe summaries.

This product shell is curated from the upstream AI Security Range repo and related public-safe outputs.

Included assets

  • • Curated lab catalog and roadmap
  • • Public metrics, scenario snapshot, Trust Scanner sample, and generation rollups
  • • Chart exports and a package README
  • • Clear runtime boundary and provenance

v1 · public-safe shell

Live · controlled environment

Atlassian Threat Canvas, LLM Attack Range, and Trust Scanner are open for threat modeling, testing, and evidence capture