genai
SECURITY LAB
Certification

GSCP — the GenAI security certification you earn against live AI targets.

Planned · releases 15 September 2026

You exploit real vulnerabilities on live AI targets and write up the fix — graded on the evidence you capture and the report you write, not a multiple-choice quiz.

Bundle includes three months of lab access. Exam alone is $499.
Certifying a team? Sittings are bought per person, on top of any plan.
The credential
GenAI Security Certified Professional
AvailabilityPlanned — 15 September 2026
FormatHands-on exam against live targets
SupervisionHuman proctor, identity verified
GradingCaptured exploit evidence + written report
ScopeOWASP LLM Top 10, Agentic AI, MCP
How you earn it

Two things have to be true to pass

The exploit has to actually work, and you have to be able to explain it. The platform proves the first objectively. A human grader judges the second.

1
Break live targets

You work against running, deliberately vulnerable AI applications — real models, real tool calls, randomized per-candidate secrets. Nothing is simulated and nothing is multiple choice.

2
The platform verifies your exploit

Each objective is graded on the evidence you capture. The harness checks what your payload actually did to the target, so a pass is a demonstrated exploit — not a claim about one.

3
Write the findings report

You submit a professional report for the vulnerabilities you land: impact, root cause, reproduction, and the remediation you would hand a product team. A grader reviews it.

4
Earn a verifiable credential

Pass both halves under proctoring and you receive a dated credential with a unique ID, publicly verifiable at its own URL.

Before any of that: work the labs, take a graded mock exam on randomized targets, then book a proctored slot.

Why now

AI shipped faster than the skills to secure it

Product teams are putting language models into production. Very few of those teams have anyone who's actually attacked one. A standard web or API pentest scope walks straight past prompt injection, agent and MCP tool abuse, and retrieval boundaries — so the gap stays open. GSCP is a dated, public record that you found, exploited and reported those flaws on running systems.

Exam composition

AI-first, but not AI-only

Real AI systems get breached through the web app and the cloud account around them, so the exam tests those too — weighted to reflect where the work actually is.

75%AI / LLM
How it differs

Earned against live targets, not recognized on a page

GSCPMultiple-choice / PDF certs
Graded onA captured exploit plus a written findings reportA multiple-choice score
You work againstLive, vulnerable AI applicationsStatic questions or a PDF
ProctoringHuman-proctored, identity verifiedOften none
You walk away withA report you can show an employerA certificate only
RetakeOne free retake, then $99 per attemptPay per attempt
Graded on
GSCP
A captured exploit plus a written findings report
Multiple-choice / PDF certs
A multiple-choice score
You work against
GSCP
Live, vulnerable AI applications
Multiple-choice / PDF certs
Static questions or a PDF
Proctoring
GSCP
Human-proctored, identity verified
Multiple-choice / PDF certs
Often none
You walk away with
GSCP
A report you can show an employer
Multiple-choice / PDF certs
A certificate only
Retake
GSCP
One free retake, then $99 per attempt
Multiple-choice / PDF certs
Pay per attempt
Syllabus

What's on the exam

Live labs behind every exam domain
OWASP LLM Top 10 2026
LLM01
Prompt injection
— labs · 6–8h

Direct overrides, jailbreaks, encoding, hidden-text smuggling, and indirect injection through retrieved content.

LLM02
Sensitive information disclosure
— labs · 3–4h

Pulling PII, credentials, and other tenants' data out of the model's context.

LLM03
Excessive agency
— labs · 4h

Over-scoped tools, unvalidated arguments, and actions that execute without a human gate.

LLM04
Supply chain
— labs · 3h

Unsigned prompts and configs, poisoned dependencies, unsafe model deserialization.

LLM05
Data and model poisoning
— labs · 2–3h

Corrupting training, fine-tuning, and ingestion pipelines.

LLM06
Unbounded consumption
— labs · 2h

Denial-of-wallet, runaway agent loops, and cost amplification.

LLM07
Misinformation
— labs · 2–3h

Inducing confident, ungrounded output and defeating citation checks.

LLM08
Hidden Context Exposure
— labs · 2h

Recovering hidden instructions and the secrets developers leave inside them.

LLM09
Vector and embedding weaknesses
— labs · 3h

Index poisoning, cross-tenant retrieval, and leakage through similarity search.

LLM10
Improper output handling
— labs · 3h

Turning model output into XSS, SSRF, and command execution in the downstream sink.

OWASP Agentic AI Top 10 2026
ASI01
Agent goal hijack
— labs · 3–4h

Redirecting an agent's objective through injected or poisoned instructions.

ASI02
Tool misuse & exploitation
— labs · 2h

Coercing agents into unsafe tool calls and abusing tool arguments.

ASI03
Identity & privilege abuse
— labs · 3–4h

Confused-deputy chains and privilege escalation across agent identities.

ASI04
Agentic supply chain
— labs · 2h

Poisoned skills, plugins, and delegated agents pulled into the loop.

ASI05
Unexpected code execution
— labs · 1–2h

Agent-driven paths that reach a real interpreter or shell.

ASI06
Memory & context poisoning
— labs · 1–2h

Planting durable instructions in an agent's memory or shared context.

ASI07
Insecure inter-agent communication
— labs · 1–2h

Trusting and relaying unvalidated messages between agents.

ASI08
Cascading failures
— labs · 1h

One compromised step propagating through a multi-agent chain.

ASI09
Human-agent trust exploitation
— labs · 1–2h

Abusing the human-in-the-loop gate to wave through unsafe actions.

OWASP MCP Top 10 2025
MCP01
Token mismanagement & secret exposure
— labs · 1–2h

Leaking or replaying the tokens and secrets an MCP server holds.

MCP02
Privilege escalation via scope creep
— labs · 1–2h

Widening a server's granted scope beyond what the task needs.

MCP03
Tool poisoning
— labs · 1–2h

Malicious tool descriptions that steer the agent into unsafe calls.

MCP04
Supply chain & dependency tampering
— labs · 1h

Compromised MCP packages, install scripts, and pinned versions.

MCP05
Command injection & execution
— labs · 1h

Reaching a shell or interpreter through an MCP tool's arguments.

MCP06
Intent flow subversion
— labs · 1–2h

Rerouting the intended sequence of tool calls mid-task.

MCP07
Insufficient authN/authZ
— labs · 1–2h

Missing or weak checks on who may call which MCP tools.

MCP08
Lack of audit & telemetry
— labs · 1h

Silent tool use with no trail to detect or reconstruct abuse.

MCP09
Shadow MCP servers
— labs · 1h

Rogue or impersonated servers registered into the client.

MCP10
Context injection & over-sharing
— labs · 1–2h

Over-broad context handed to a server, and injection back through it.

Counts are live from the catalogue. A lab can cover more than one category, so the columns overlap and sum above the distinct total shown above.

OWASP LLM Top 10
Native — the twelve exam domains are its ten categories plus agentic and MCP
MITRE ATLAS
Offensive scenarios reference real ATLAS adversary techniques
NIST AI RMF
Scenarios map to the Govern, Map, Measure and Manage functions
Verifiable at a public URL
What you receive

A credential anyone can check

Issued on passing, scoped to the domains you were assessed against, and verifiable by anyone at its public URL.

Sample — not an issued credential
genai
SECURITY LAB
Professional Certification
This is to certify that
Sample Candidate
GenAI Security Certified Professional (GSCP)
Practical exam · live exploit + remediation · proctored · OWASP LLM, Agentic & MCP
Certified
Assessed across the OWASP LLM, Agentic & MCP Top 10s
OWASP LLM Top 10
LLM01–10
OWASP Agentic Top 10
ASI01–10
OWASP MCP Top 10
MCP01–10
Aditya Agrawal
Aditya Agrawal
Founder, GenAI Security Lab
GENAI SECURITY LABCERTIFIED PROFESSIONALgenai
Issued
15 Jul 2026
Valid through
15 Jul 2029 · recertify to renew
Credential ID
GSCP-2026-04871
genaisecuritylab.com/verify
Scan to verify
Every GSCP resolves to a public record
genaisecuritylab.com/verify/GSCP-2026-04871
Is GSCP for you?

Sit it now, or train first?

This exam assumes you can already test. Answer three things and we'll tell you whether to book it or spend time on the labs first.

Your background
What you have already done
Where AI sits in your work
Select every answer that applies

We will point you at the exam or the training path based on your testing, reporting, and LLM experience.

Questions

Before you register

A human proctor verifies your identity and supervises the session from start to finish. That supervision is what lets an employer treat the result as yours.

Prove it on a live target.

The bundle gives you three months of labs to prepare and one proctored attempt at the credential.