HackIndex.io
Language
AI and machine learning systems introduce a new class of attack surfaces to pentesting. As organizations deploy large language models (LLMs), ML inference APIs, and AI-powered applications, these systems become targets for prompt injection, model manipulation, and data extraction. AI security testing requires understanding both traditional application security and the unique behaviors of ML models.
AI-specific attacks focus on prompt injection (direct and indirect), jailbreaking model safety filters, training data extraction, model inversion, adversarial inputs that cause misclassification, and supply chain risks from poisoned models or compromised model repositories. LLM-integrated applications are particularly vulnerable where user input reaches the model alongside system prompts or retrieval-augmented generation (RAG) context.
This section covers AI security testing techniques: enumeration of model capabilities, system prompts, and tool integrations, exploitation through prompt injection to bypass access controls or exfiltrate data, manipulation of AI-powered decision systems, and assessment of ML supply chain risks including model provenance and dependency poisoning.
Guides for mapping AI endpoints, models, APIs, and dependencies
Techniques for identifying exposed AI/LLM endpoints, APIs, and model interfaces during recon, including fingerprinting frameworks, discovering hidden routes, and mapping AI attack surfaces.
Techniques for discovering and mapping vector database instances, embedding endpoints, and RAG pipeline components to identify attack surface in AI-driven applications.
Techniques for mapping AI supply chain components including models, datasets, APIs, and dependencies to identify attack surfaces in AI platform environments.
Techniques to identify AI model providers, versions, and architectures through behavioral probing, response analysis, and metadata extraction.
Techniques for discovering publicly exposed ML model servers, inference endpoints, and AI APIs through passive recon, port scanning, and service fingerprinting.
Identify and enumerate AI/ML frameworks, SDKs, and third-party dependencies within target environments to uncover attack surface and supply chain risks.
Guides on mapping AI trust boundaries, agent permissions, and targets
Techniques for identifying critical assets in AI systems, including models, training data, APIs, and inference pipelines to prioritize security controls.
Identifies SPOF risks in multi-agent AI architectures, covering orchestrator compromise, shared memory poisoning, and cascading failure vectors across agent trust boundaries.
Explores trust boundary identification in AI system architectures, covering data flows, model endpoints, and inter-component trust zones for effective threat modeling.
Analyzes agent-to-agent communication patterns, trust boundaries, and authentication mechanisms to identify attack surfaces in multi-agent AI architectures.
Analyze and document AI agent permission boundaries, tool access scopes, and memory persistence to identify privilege escalation and data leakage risks.
Guides covering vuln discovery in LLM, RAG, MCP, and AI infra
Identifies security flaws in AI agent tool integrations, covering authentication gaps, input validation failures, and unsafe API connections that expose systems to exploitation.
Identify prompt injection vulnerabilities in AI systems by probing instruction boundaries, testing context isolation, and detecting system prompt leakage or override weaknesses.
Identifies vulnerabilities in AI supply chains including model dependencies, training pipelines, third-party components, and infrastructure misconfigurations across AI platforms.
Identify misconfigured AI inference endpoints exposing unauthenticated or over-permissioned access through API probing, header analysis, and permission enumeration techniques.
Techniques for discovering exposed Model Context Protocol endpoints, authentication gaps, and misconfigured AI platform attack surfaces during reconnaissance.
Techniques for identifying attack surfaces in RAG pipelines, including document ingestion vectors, embedding manipulation, and retrieval mechanism abuse for security assessment.
Exploit LLM apps, RAG pipelines, agents, and AI infrastructure
Exploiting embedding spaces and model internals to extract sensitive information, reverse-engineer training data, and manipulate AI platform behaviour.
Craft adversarial inputs to manipulate AI model outputs, exploiting gradient-based perturbations and transferability to bypass classifiers and deceive neural networks.
Techniques to invert ML embeddings and reconstruct sensitive training data, exposing PII and proprietary information from compromised AI platform models.
Demonstrates how adversaries query AI models to determine if specific data was used in training, exploiting membership inference to leak sensitive training set information.
Systematically query AI models to reverse-engineer architecture, training data, and decision boundaries through crafted input enumeration and response analysis.
Techniques for extracting sensitive training data, system prompts, or confidential context from AI models using adversarial prompt injection and jailbreaking methods.
Guides on persistence, pivoting, and data extraction from AI systems
Techniques for discovering and mapping tools, APIs, and external services connected to compromised AI platforms to expand attack surface visibility.
Techniques for extracting hidden system prompts and proprietary instructions from deployed AI models using prompt injection, jailbreaks, and inference attacks.
Techniques for extracting stored conversation histories, user profiles, and session data from compromised AI platforms during post-exploitation.
Techniques for maintaining persistent LLM manipulation by injecting malicious entries into RAG vector stores, corrupting retrieval pipelines post-compromise.
Techniques for laterally moving from compromised AI infrastructure into traditional network segments, exploiting trust relationships, misconfigurations, and shared credentials.
Guides on poisoning datasets, weights, adapters, and pipelines
Exploiting vulnerabilities in ML fine-tuning workflows to inject malicious weights, backdoors, or poisoned datasets into AI model supply chains.
Explores how adversaries inject malicious training data into AI pipelines to manipulate model outputs, backdoor behavior, or degrade performance at inference.
Explores dependency confusion attacks targeting AI Python packages, covering namespace hijacking, private package spoofing, and supply chain poisoning mitigations.
Explores backdoor injection via malicious LoRA adapters and fine-tuning layers, targeting AI model supply chains to compromise downstream deployments.
Covers risks of compromised model weights distributed via public registries, detailing detection, integrity verification, and mitigation strategies for AI supply chain threats.
We use cookies to improve your browsing experience, analyze site traffic, and personalize content. By clicking "Accept All", you consent to our use of cookies. Privacy Policy.