412
techniques cataloged
47
new this month
3
critical active

Your LLMs are vulnerable.
Know every attack.
Defend every endpoint.

The structured attack catalog, defensive testing suite, and hands-on training labs that LLM security professionals actually need. Think MITRE ATT&CK meets HackTheBox.

Join 1,847 security professionals already on the list

412 techniques8 categories47 new this month3 critical active

THE PLATFORM

Catalog. Defend. Train.

Three layers that compound: know the attacks, test your defenses, train your team.

CRITICAL

Attack Catalog

412 classified jailbreak techniques with severity scoring, affected models, detection signatures, and mitigations. Structured like MITRE ATT&CK, updated weekly.

SCAN

Defender

Point at your LLM API endpoint. Run curated attack suites. Get a security report with specific vulnerabilities and mitigations. PDF, JSON, and SARIF export.

LEARN

Academy

Hands-on labs in sandboxed environments. Attack labs, defense labs, incident response scenarios. Progressive difficulty. Verifiable credentials.

ATTACK CATALOG

Not just a list — a living threat intelligence database

Every technique classified by severity, affected models, detection signatures, and proven mitigations. The catalog is structured, searchable, and cross-referenced — not a blog post that goes stale.

Full reproduction steps for authenticated users with verified security credentials. Detection signatures and mitigations are public — defense-first for everyone.

8 attack categories with expandable subcategories
Severity scoring (CVSS-LLM) with affected model matrix
Detection signatures you can ship to production
Mitigations with code examples, not just prose
Cross-references between related techniques
AJS-2026-0091CRITICAL

Tool Parameter Injection

When an LLM has access to tools with parameterized inputs, user-supplied content can be passed directly to tool parameters without sanitization. Enables classical injection attacks via the LLM as intermediary.

Affected Models

· GPT-4oConfirmed
· Claude 3.5Confirmed
· Gemini 2.0Confirmed

Detection Signatures

· Tool calls containing SQL/shell metacharacters

· User message → tool param content overlap

· Anomalous tool parameter length/entropy

Mitigations

· Parameterized queries for all tool inputs

· Tool input validation layer (type + format)

· Output monitoring for injection indicators

DEFENDER

Scan your endpoints. Fix what's broken.

Point the Defender at your LLM API. It runs curated attack suites — not raw prompt spam, but intelligent, staged tests modeled on real-world jailbreak techniques. Every finding links back to the catalog with specific mitigations.

TARGET: api.example.com/v1/chatMODEL: GPT-4o (detected)TESTED: 247 / 412

Results

CRITICAL
3
HIGH
7
MEDIUM
14
LOW
22
PASS
201
PDF ReportJSONSARIF

HOW IT WORKS

Three steps to LLM security coverage

01

Browse the catalog

Explore 412 classified techniques by category, model, or severity. Understand the attack surface before you test it.

02

Scan your endpoints

Run curated attack suites against your LLM API. Get actionable findings with specific mitigations. CI/CD integration available.

03

Train your team

Hands-on labs in sandboxed environments. Your team learns by attacking and defending real LLM deployments. Earn verifiable credentials.

FAQ

Common questions

Is this legal?

Yes. The Defender tool runs against endpoints you own and authorize. Catalog techniques are published under responsible disclosure. We follow CVE-equivalent processes.

Do you publish working jailbreaks?

Detection signatures and mitigations are public. Full reproduction steps require authenticated access with verified security professional credentials.

What models does this cover?

All major commercial LLMs (GPT-4, Claude, Gemini, Llama, Mistral) plus model-agnostic techniques. The catalog is model-aware — filter by what you deploy.

How is this different from OWASP LLM Top 10?

OWASP gives you 10 risk categories. We give you 412 specific techniques with reproduction steps, detection signatures, and hands-on labs. It’s the difference between knowing SQL injection is a risk and knowing exactly how to find and fix it.

Can I use this for my team?

Team plans include shared dashboards, assigned training paths, and Defender scans for your production endpoints. Enterprise plans add SSO and self-hosted scanning.

How often is the catalog updated?

Weekly. New techniques are indexed from research papers, responsible disclosures, and community submissions. Monthly ‘State of Jailbreaking’ report tracks trends.

Your LLMs are exposed.
Let's fix that.

JailbreakingSite is building the security resource LLM teams actually need. Get early access and shape the platform.

Founding members get lifetime access to the Pro catalog