AREA 01
Agentic AI security & tool-use authorization
Confine a tool-using agent: find and prove authorization failures, confused-deputy paths and unsafe tool chaining, then establish boundaries that hold under re-test.
RANGER PARALLAX
AI security, adversarial systems and digital trust.
An advanced independent technical qualification pathway. PARALLAX challenges experienced practitioners to establish trust, identify material weaknesses, demonstrate real technical consequences, engineer defenses, and independently verify outcomes across intelligent systems. It is not a deepfake-detection course or an introduction to artificial intelligence.
Registering interest is not an application or admission. It does not create a candidate account, guarantee a cohort seat, or grant lab access.
PathwayIndependent of STRATA
StructureIntegrated challenge areas
QualificationSeparately governed
ApplicationsNot yet open
evidence modelEd25519 signed
What PARALLAX is
A qualification for intelligent systems, not a course about them.
PARALLAX is a separate pathway from STRATA, with its own challenge architecture, evidence model, and separately governed standard. It is built around integrated challenge areas rather than a fixed board sequence. Practitioners establish trust in AI systems, prove material weaknesses with real technical consequences, engineer and verify defenses, and defend every conclusion with reproducible, signed evidence. It does not inherit STRATA's six-board sequence, 26-week schedule, or 95% standard.
Challenge areas
Six integrated competency domains.
PARALLAX challenges span the lifecycle of intelligent systems. They are integrated rather than sequential: a single engagement can cross several domains, and each requires independent investigation and defensible evidence.
AREA 02
LLM & RAG security
Find and prove trust-boundary and data-isolation failures in an LLM or RAG application, including direct and indirect prompt injection and cross-tenant leakage, then remediate verifiably.
AREA 03
Adversarial AI evaluation & mitigation validation
Run a structured adversarial evaluation, characterize failure modes with reproducible evidence, and validate that mitigations actually hold on held-out probes.
AREA 04
Model & AI supply-chain integrity
Investigate model and artifact integrity: tampering, unsafe deserialization and broken provenance, and design controls that establish and prove integrity.
AREA 05
Synthetic-media forensics & multimodal fraud
Conduct a defensible forensic investigation of synthetic and real media, reaching a calibrated determination supported by reproducible observations and verified provenance.
AREA 06
AI monitoring, incident response & assurance
Detect, triage and respond to an AI-system incident from telemetry, then assure and recover, with evidence that recomputes.
Proposed technical environments
Environments are proposed, not yet operational.
PARALLAX engagements run against Ranger-owned, vulnerable-by-design intelligent systems, isolated per engagement. These environments are in design. None are deployed, and no candidate access exists.
ENVproposed
Agents & tool APIs
Tool-using agents and protected tool endpoints with authorization boundaries to defeat and prove.
ENVproposed
LLM & RAG applications
Retrieval and generation stacks with seeded trust-boundary, data-isolation and injection defects.
ENVproposed
Media & provenance
Synthetic and authentic media sets for calibrated forensic determination and provenance verification.
Environments, inference systems, assessment criteria and qualifications described here are in development. They are not operational, and nothing on this page grants access.
Technology approach
Provider-independent, chosen by challenge.
PARALLAX is provider-independent. The evaluation ecosystem is chosen by benchmark and by what each challenge actually requires, not by vendor, and no component is adopted to claim breadth. Each integration is tracked as proposed, validated or operational. Hugging Face is one prospective inference integration, not the technology that defines the pathway. Protected scenarios, evaluator internals, private prompts and infrastructure credentials are never exposed.
Model inference
Hugging Face and alternative hosted-inference providers; candidate models such as Qwen and gpt-oss selected by benchmark. No single provider defines the pathway.
Proposed
Adversarial AI evaluation
Microsoft PyRIT and NVIDIA garak for structured adversarial probing and LLM vulnerability assessment.
Proposed
Application, agent & RAG testing
Promptfoo for application, agent and retrieval testing; Inspect AI for reproducible evaluation workflows.
Proposed
Target systems under test
Ranger-owned vulnerable-by-design AI applications, agents, retrieval systems and protected tool APIs, built and held by Ranger.
Proposed
Media & provenance forensics
FFmpeg, ExifTool, OpenCV and C2PA for synthetic-media and multimodal-fraud investigation.
Proposed
Investigation & signed evidence
OpenSearch and the Ranger Ed25519 signed-evidence system, the shared platform plane PARALLAX will reuse.
Operational
Proposed
Validated
Operational
Status reflects integration maturity on the platform, not candidate progress or any vendor certification. PyRIT, garak, Promptfoo, Inspect AI and other frameworks are candidate tooling, not deployed assessment components.
Verification philosophy
Trust is demonstrated, then independently verified.
PARALLAX shares the platform's evidence discipline. Qualification is never decided by AI. Every outcome must recompute from recorded inputs and is sealed with an Ed25519 signature that can be re-verified independently.
- Reproducible evidence: findings recompute from recorded inputs, not screenshots
- Held-out validation: mitigations are tested against probes the candidate has not seen
- Independent re-measurement: outcomes are re-measured and signed, not self-reported
- Protected truth: scenario truth, evaluator logic and private prompts are never exposed
candidate-facing inferencenone
qualification decisiondeterministic · signed
assessment componentsin design
candidate accessnone
training on candidate datadisabled
No inference system or assessment is claimed operational until it is deployed and independently verified.
Register interest
Submit interest in PARALLAX.
PARALLAX is in development. Register interest to be contacted when the pathway opens. This is not an application and is kept separate from STRATA admission.
- No candidate account is created and no cohort is assigned
- No lab authorization and no access is granted
- We contact you only about PARALLAX, and only with your consent
Applying for a STRATA cohort instead? Apply for consideration.