Cinder Security — AI Red Team as a Service

● GHSA-m4rw-22q2-87j8 — ModelEngine SSRF via Prompt Injection — CVE Pending ● GHSA-4fpw-hjmg-x4qr — LangGraph RAG Poisoning — CVSS 7.6 ● 2 GitHub Security Advisories Published in 2026 ● Responsible Disclosure — All Findings Reported to Vendors ● Fracture — Autonomous AI Red Team Engine — MIT License ● OWASP LLM Top 10 Full Coverage ● GHSA-m4rw-22q2-87j8 — ModelEngine SSRF via Prompt Injection — CVE Pending ● GHSA-4fpw-hjmg-x4qr — LangGraph RAG Poisoning — CVSS 7.6 ● 2 GitHub Security Advisories Published in 2026 ● Responsible Disclosure — All Findings Reported to Vendors ● Fracture — Autonomous AI Red Team Engine — MIT License ● OWASP LLM Top 10 Full Coverage

// The Problem

AI is everywhere.
AI security is nowhere.

Organizations are deploying AI at unprecedented speed. But most have zero offensive security testing for their AI systems. The attack surface is massive and growing.

88%

Jailbreak success rate on major LLMs via psychological manipulation (HPM)

97%

Attack success rate via fine-tuning backdoors on GPT-4 class models

<5%

Of companies deploying AI have done any offensive security testing

Cost to fully compromise a model through fine-tuning API attacks

GitHub Security Advisories published by Cinder Security in 2026

// Impact by the Numbers

Research that ships.

Every number below is a real finding, a real vendor notified, a real patch shipped.

GitHub Security Advisories Published

2026

Active Vulnerability Research Cases

CSR-2026-*

Vendors Notified Under Responsible Disclosure

2026

Critical Patches Shipped by Vendors

Confirmed fix live

OWASP LLM Top 10 Vectors Covered

Full coverage

// Research Findings

Vulnerabilities we've disclosed.

Real-world findings from active research on production AI systems. All responsibly disclosed to vendors. All with reproducible proof-of-concepts.

CSR-2026-002

ModelEngine / fit-framework

SSRF via Prompt Injection — LangChain RequestsGetTool(allow_dangerous_requests=True) with no URL filtering enables cloud metadata exfiltration via IAM token theft on AWS and Alibaba Cloud through 169.254.169.254.

Critical ✓ Patch live — CVE pending

GHSA-m4rw-22q2-87j8 ↗

CSR-2026-007

LangGraph / LangChain

Indirect prompt injection via RAG poisoning — a single poisoned document hijacks a ReAct agent's tool calls, enabling persistent instruction injection across all subsequent interactions with no code execution required.

High — CVSS 7.6 ✓ Public advisory

GHSA-4fpw-hjmg-x4qr ↗

CSR-2026-003

Insightify

Azure OpenAI API credentials exposed in plaintext config + allow_dangerous_code=True — enabling API key theft and arbitrary code execution on the host server via a single prompt.

Critical ⏳ Disclosure in progress

Active Research

Multiple AI Frameworks

Ongoing vulnerability research across AI agent frameworks, guardrail systems, and RAG pipelines. Coordinated disclosure in progress with multiple vendors. Additional advisories to be published upon confirmation.

In Progress ⏳ Coordinated disclosure

// Public Advisories

Security Advisories.

Published GitHub Security Advisories and CVEs from Cinder Security research.

Advisory	CSR ID	Target	Type	CVSS	Status
GHSA-m4rw-22q2-87j8	CSR-2026-002	ModelEngine fit-framework	SSRF + Prompt Injection	Critical	Patch Live
GHSA-4fpw-hjmg-x4qr	CSR-2026-007	LangGraph / LangChain	RAG Poisoning	7.6 High	Public
CVE pending	CSR-2026-002	ModelEngine fit-framework	CVE Assignment	Critical	MITRE Pending

// Open Source Tool

MIT License

Fracture — autonomous AI red team engine.

An open-source framework for automated offensive testing of AI systems. Fracture runs structured attack campaigns against LLMs and AI agents — fingerprinting, extracting system prompts, poisoning memory, and escalating privileges through multi-turn psychological manipulation.

cinder-security/fracture

MODULES

● fingerprint

● extract

● memory

● hpm

● ssrf

● retrieval_poison

● auto

v0.1.0 · Python · MIT

FRACTURE — AUTONOMOUS AI RED TEAM ENGINE

// Services

Full-spectrum AI offensive security.

We don't just scan — we think like attackers. Every engagement is tailored to your specific AI stack and threat model.

⚔️

AI Penetration Testing

Comprehensive one-time security assessment of your AI systems. We test every attack vector and deliver a detailed report with reproducible proof-of-concepts, CVSS scores, and remediation guidance.

One-time engagement

🔄

Continuous AI Red Teaming

Ongoing offensive testing as your AI evolves. Every model update, every new feature — we test it before your users find the gaps. Monthly retainer with prioritized findings.

Monthly retainer

🎓

AI Security Training

Hands-on workshops for your engineering and security teams. Learn to think like an AI attacker and build more resilient systems from day one.

Workshop

// Pricing

Transparent pricing.
No surprises.

Straightforward engagements with clear deliverables. Every assessment includes a professional PDF report and a 30-minute debrief call.

Entry

$750

USD — one-time

AI Security Assessment

→ Up to 3 attack vectors tested
→ Reproducible PoC per finding
→ Professional PDF report
→ 30-min debrief call
→ Delivered in 5 business days
→ 50% upfront · 50% on delivery

Get Started

What we test.

The full attack surface of modern AI systems — from prompt-level exploits to infrastructure-level compromises.

Direct & Indirect Prompt Injection

Multi-turn Jailbreak Attacks

Psychological Manipulation (HPM)

System Prompt Extraction

RAG Pipeline Poisoning

Fine-tuning Backdoors & Data Poisoning

SSRF via AI Agent Tool Abuse

Tool & Function Call Abuse

Multi-Agent Attack Chains

Model & API Key Extraction

Memory Poisoning in Persistent Agents

Guardrail & Filter Bypass

// Process

How we work.

A structured approach to finding what others miss.

Scope & Profile

Map your AI stack, identify attack surfaces, and define engagement rules.

Attack & Exploit

Execute targeted attacks across all vectors. Every finding includes a reproducible PoC.

Report & Remediate

Detailed security report with severity ratings, CVSS scores, and fix recommendations.

Verify & Harden

Re-test after fixes. Confirm vulnerabilities are resolved and defenses hold.

We break AI systems
before attackers do.

AI is everywhere.
AI security is nowhere.

Research that ships.

Vulnerabilities we've disclosed.

Security Advisories.

Fracture — autonomous AI red team engine.

Full-spectrum AI offensive security.

AI Penetration Testing

Continuous AI Red Teaming

AI Security Training

Transparent pricing.
No surprises.

What we test.

How we work.

Scope & Profile

Attack & Exploit

Report & Remediate

Verify & Harden

Ready to find out what's burning?

We break AI systemsbefore attackers do.

AI is everywhere.AI security is nowhere.

Research that ships.

Vulnerabilities we've disclosed.

Security Advisories.

Fracture — autonomous AI red team engine.

Full-spectrum AI offensive security.

AI Penetration Testing

Continuous AI Red Teaming

AI Security Training

Transparent pricing.No surprises.

What we test.

How we work.

Scope & Profile

Attack & Exploit

Report & Remediate

Verify & Harden

Ready to find out what's burning?

We break AI systems
before attackers do.

AI is everywhere.
AI security is nowhere.

Transparent pricing.
No surprises.