AI Red Teaming Roadmap
This AI Red Teaming Roadmap is a free, full-text syllabus of 64 lessons in 5 stages, starting with Introduction and ending with Infrastructure Security. It suits anyone who wants to learn AI Red Teaming, whether new to it or filling gaps. There is no fixed length: every lesson is open from the start, so you go at your own pace and skip what you already know.
Last updated
Reading is open to everyone. Sign in to join this roadmap and track what you finish.
Continue learning
1. IntroductionIntroduction
0%Introduction
5 topics · 4 sub-topics · 0 completed
The diagram draws the same 64 lessons as a map, with lines showing the order. It needs JavaScript; the full syllabus below lists every lesson as text.
Full syllabus: every lesson in the AI Red Teaming roadmap
5 stages and 64 lessons, in the order the roadmap suggests. Every lesson is free to read and open from the start.
Stage 1: Introduction
Introduction
AI Red Teaming is the practice of simulating adversarial attacks against AI systems to proactively identify vulnerabilities, potential misuse scenarios, and failure modes before malicious actors do. Distinct from traditional cybersecurity red teaming, it focuses on the unique attack surfaces of AI models, such as…
AI Security Fundamentals
This covers the foundational concepts essential for AI Red Teaming, bridging traditional cybersecurity with AI-specific threats. An AI Red Teamer must understand common vulnerabilities in ML models (like evasion or poisoning), security risks in the AI lifecycle (from data collection to deployment), and how AI…
Why Red Team AI Systems?
AI systems introduce novel risks beyond traditional software, such as emergent unintended capabilities, complex failure modes, susceptibility to subtle data manipulations, and potential for large-scale misuse (e.g., generating disinformation). AI Red Teaming is necessary because standard testing methods often fail…
Ethical Considerations
Ethical conduct is crucial for AI Red Teamers. While simulating attacks, they must operate within strict legal and ethical boundaries defined by rules of engagement, focusing on improving safety without causing real harm or enabling misuse. This includes respecting data privacy, obtaining consent where necessary,…
Role of Red Teams
The role of an AI Red Team is to rigorously challenge AI systems from an adversarial perspective. They design and execute tests to uncover vulnerabilities related to the model's logic, data dependencies, prompt interfaces, safety alignments, and interactions with surrounding infrastructure. They provide detailed…
Stage 2: Prompt Hacking
Prompt Hacking
Prompt hacking is a core technique for AI Red Teamers targeting LLMs. It involves crafting inputs (prompts) to manipulate the model into bypassing safety controls, revealing hidden information, or performing unintended actions. Red teamers systematically test various prompt hacking methods (like jailbreaking,…
Supervised Learning
AI Red Teamers analyze systems built using supervised learning to probe for vulnerabilities like susceptibility to adversarial examples designed to cause misclassification, sensitivity to data distribution shifts, or potential for data leakage related to the labeled training data. Understanding how these models…
Unsupervised Learning
When red teaming AI systems using unsupervised learning (e.g., clustering algorithms), focus areas include assessing whether the discovered patterns reveal sensitive information, if the model can be manipulated to group data incorrectly, or if dimensionality reduction techniques obscure security-relevant features.…
Reinforcement Learning
Red teaming RL-based AI systems involves testing for vulnerabilities such as reward hacking (exploiting the reward function to induce unintended behavior), unsafe exploration (agent takes harmful actions during learning), or susceptibility to adversarial perturbations in the environment's state. Understanding the…
Neural Networks
Understanding neural network architectures (layers, nodes, activation functions) is vital for AI Red Teamers. This knowledge allows for targeted testing, such as crafting adversarial examples that exploit specific activation functions or identifying potential vulnerabilities related to network depth or…
Generative Models
AI Red Teamers focus heavily on generative models (like GANs and LLMs) due to their widespread use and unique risks. Understanding how they generate content is key to testing for issues like generating harmful/biased outputs, deepfakes, prompt injection vulnerabilities, or leaking sensitive information from their…
Large Language Models
LLMs are a primary target for AI Red Teaming. Understanding their architecture (often Transformer-based), training processes (pre-training, fine-tuning), and capabilities (text generation, summarization, Q&A) is essential for identifying vulnerabilities like prompt injection, jailbreaking, data regurgitation, and…
Prompt Engineering
For AI Red Teamers, prompt engineering is both a tool and a target. It's a tool for crafting inputs to test model boundaries and vulnerabilities (e.g., creating jailbreak prompts). It's a target because understanding how prompts influence LLMs is key to identifying prompt injection vulnerabilities and designing…
Confidentiality, Integrity, Availability
The CIA Triad is directly applicable in AI Red Teaming. Confidentiality tests focus on preventing leakage of training data or proprietary model details. Integrity tests probe for susceptibility to data poisoning or model manipulation. Availability tests assess resilience against denial-of-service attacks targeting…
Threat Modeling
AI Red Teams apply threat modeling to identify unique attack surfaces in AI systems, such as manipulating training data, exploiting prompt interfaces, attacking the model inference process, or compromising connected tools/APIs. Before attacking an AI system, red teamers perform threat modeling to map out possible…
Risk Management
AI Red Teamers contribute to the AI risk management process by identifying and demonstrating concrete vulnerabilities. Findings from red team exercises inform risk assessments, helping organizations understand the likelihood and potential impact of specific AI threats and prioritize resources for mitigation based…
Vulnerability Assessment
While general vulnerability assessment scans infrastructure, AI Red Teaming extends this to assess vulnerabilities specific to the AI model and its unique interactions. This includes probing for prompt injection flaws, testing for adversarial example robustness, checking for data privacy leaks, and evaluating…
Stage 3: Model Vulnerabilities
Model Vulnerabilities
This category covers attacks and tests targeting the AI model itself, beyond the prompt interface. AI Red Teamers investigate inherent weaknesses in the model's architecture, training data artifacts, or prediction mechanisms, such as susceptibility to data extraction, poisoning, or adversarial manipulation.
Model Weight Stealing
AI Red Teamers assess the risk of attackers reconstructing or stealing the proprietary weights of a trained model, often through API query-based attacks. Testing involves simulating such attacks to understand how easily the model's functionality can be replicated, which informs defenses like query rate limiting,…
Unauthorized Access
AI Red Teamers test if vulnerabilities in the AI system or its interfaces allow attackers to gain unauthorized access to data, functionalities, or underlying infrastructure. This includes attempting privilege escalation via prompts, exploiting insecure API endpoints connected to the AI, or manipulating the AI to…
Stage 4: Code Injection
Code Injection
AI Red Teamers test for code injection vulnerabilities specifically in the context of AI applications. This involves probing whether user input, potentially manipulated via prompts, can lead to the execution of unintended code (e.g., SQL, OS commands, or script execution via generated code) within the application…
Jailbreak Techniques
Jailbreaking is a specific category of prompt hacking where the AI Red Teamer aims to bypass the LLM's safety and alignment training. They use techniques like creating fictional scenarios, asking the model to simulate an unrestricted AI, or using complex instructions to trick the model into generating content that…
Safety Filter Bypasses
AI Red Teamers specifically target the safety mechanisms (filters, guardrails) implemented within or around an AI model. They test techniques like using synonyms for blocked words, employing different languages, embedding harmful requests within harmless text, or using character-level obfuscation to evade detection…
Prompt Injection
Prompt injection is a critical vulnerability tested by AI Red Teamers. They attempt to insert instructions into the LLM's input that override its intended system prompt or task, causing it to perform unauthorized actions, leak data, or generate malicious output. This tests the model's ability to distinguish trusted…
Direct
Direct injection attacks occur when malicious instructions are inserted directly into the prompt input field by the user interacting with the LLM. AI Red Teamers use this technique to assess if basic instructions like "Ignore previous prompt" can immediately compromise the model's safety or intended function,…
Indirect
Indirect injection involves embedding malicious prompts within external data sources that the LLM processes, such as websites, documents, or emails. AI Red Teamers test this by poisoning data sources the AI might interact with (e.g., adding hidden instructions to a webpage summarized by the AI) to see if the AI…
Countermeasures
AI Red Teamers must also understand and test defenses against prompt hacking. This includes evaluating the effectiveness of input sanitization, output filtering, instruction demarcation (e.g., XML tagging), contextual awareness checks, model fine-tuning for resistance, and applying the principle of least privilege…
Data Poisoning
AI Red Teamers simulate data poisoning attacks by evaluating how introducing manipulated or mislabeled data into potential training or fine-tuning datasets could compromise the model. They assess the impact on model accuracy, fairness, or the potential creation of exploitable backdoors, informing defenses around…
Adversarial Examples
A core AI Red Teaming activity involves generating adversarial examples – inputs slightly perturbed to cause misclassification or bypass safety filters – to test model robustness. Red teamers use various techniques (gradient-based, optimization-based, or black-box methods) to find inputs that exploit model…
Model Inversion
AI Red Teamers perform model inversion tests to assess if an attacker can reconstruct sensitive training data (like images, text snippets, or personal attributes) by repeatedly querying the model and analyzing its outputs. Success indicates privacy risks due to data memorization, requiring mitigation techniques…
Stage 5: Infrastructure Security
Infrastructure Security
AI Red Teamers assess the security posture of the infrastructure hosting AI models (cloud environments, servers, containers). They look for misconfigurations, unpatched systems, insecure network setups, or inadequate access controls that could allow compromise of the AI system or leakage of sensitive data/models.
Insecure Deserialization
AI Red Teamers investigate if serialized objects used by the AI system (e.g., for saving model states, configurations, or transmitting data) can be manipulated by an attacker. They test if crafting malicious serialized objects could lead to remote code execution or other exploits when the application deserializes…
Remote Code Execution
AI Red Teamers attempt to achieve RCE on systems hosting or interacting with AI models. This could involve exploiting vulnerabilities in the AI framework itself, the web server, connected APIs, or tricking an AI agent with code execution capabilities into running malicious commands provided via prompts. RCE is…
Adversarial Training
AI Red Teamers evaluate the effectiveness of adversarial training as a defense. They test if models trained on adversarial examples are truly robust or if new, unseen adversarial attacks can still bypass the hardened defenses. This helps refine the adversarial training process itself.
Robust Model Design
AI Red Teamers assess whether choices made during model design (architecture selection, regularization techniques, ensemble methods) effectively contribute to robustness against anticipated attacks. They test if these design choices actually prevent common failure modes identified during threat modeling.
Continuous Monitoring
AI Red Teamers assess the effectiveness of continuous monitoring systems by attempting attacks and observing if detection mechanisms trigger appropriate alerts and responses. They test if monitoring covers AI-specific anomalies (like sudden shifts in output toxicity or unexpected resource consumption by the model)…
Black Box Testing
In AI Red Teaming, black-box testing involves probing the AI system with inputs and observing outputs without any knowledge of the model's architecture, training data, or internal logic. This simulates an external attacker and is crucial for finding vulnerabilities exploitable through publicly accessible…
White Box Testing
White-box testing in AI Red Teaming grants the tester full access to the model's internals (architecture, weights, training data, source code). This allows for highly targeted attacks, such as crafting precise adversarial examples using gradients, analyzing code for vulnerabilities, or directly examining training…
Grey Box Testing
Grey-box AI Red Teaming involves testing with partial knowledge of the system, such as knowing the model type (e.g., GPT-4), having access to some documentation, or understanding the general system architecture but not having full model weights or source code. This allows for more targeted testing than black-box…
Automated vs Manual
AI Red Teaming typically employs a blend of automated tools (for large-scale scanning, fuzzing prompts, generating basic adversarial examples) and manual human testing (for creative jailbreaking, complex multi-stage attacks, evaluating nuanced safety issues like bias). Automation provides scale, while manual…
Continuous Testing
Applying continuous testing principles to AI security involves integrating automated red teaming checks into the development pipeline (CI/CD). This allows for regular, automated assessment of model safety, robustness, and alignment as the model or application code evolves, catching regressions or new…
API Protection
AI Red Teamers rigorously test the security of APIs providing access to AI models. They probe for OWASP API Top 10 vulnerabilities like broken authentication/authorization, injection flaws, security misconfigurations, and lack of rate limiting, specifically evaluating how these could lead to misuse or compromise of…
Authentication
AI Red Teamers test the authentication mechanisms controlling access to AI systems and APIs. They attempt to bypass logins, steal or replay API keys/tokens, exploit weak password policies, or find flaws in MFA implementations to gain unauthorized access to the AI model or its management interfaces.
Authorization
AI Red Teamers test authorization controls to ensure that authenticated users can only access the AI features and data permitted by their roles/permissions. They attempt privilege escalation, try to access other users' data via the AI, or manipulate the AI to perform actions beyond its authorized scope.
Conferences
Attending major cybersecurity conferences (DEF CON, Black Hat, RSA) and increasingly specialized AI Safety/Security conferences allows AI Red Teamers to learn about cutting-edge research, network with peers, and discover new tools and attack/defense techniques.
Research Groups
Following and potentially contributing to research groups at universities (like CMU, Stanford, Oxford), non-profits (like OpenAI, Anthropic), or government bodies (like UK's AISI) focused on AI safety, security, and alignment provides deep insights into emerging threats and mitigation strategies relevant to AI Red…
Forums
Engaging in online forums, mailing lists, Discord servers, or subreddits dedicated to AI security, adversarial ML, prompt engineering, or general cybersecurity helps AI Red Teamers exchange knowledge, ask questions, learn about new tools/techniques, and find collaboration opportunities.
Testing Platforms
Platforms used by AI Red Teamers range from general penetration testing OS distributions like Kali Linux to specific AI red teaming tools/frameworks like Microsoft's PyRIT or Promptfoo, and vulnerability scanners like OWASP ZAP adapted for API testing of AI services. These platforms provide the toolsets needed to…
Monitoring Solutions
AI Red Teamers interact with monitoring tools primarily to test their effectiveness (evasion) or potentially exploit vulnerabilities within them. Understanding tools like IDS (Snort, Suricata), network analyzers (Wireshark), and SIEMs helps red teamers simulate attacks that might bypass or target these defensive…
Benchmark Datasets
AI Red Teamers may use or contribute to benchmark datasets specifically designed to evaluate AI security. These datasets (like HackAprompt, SecBench, NYU CTF Bench, CySecBench) contain prompts or scenarios targeting vulnerabilities, safety issues, or specific cybersecurity capabilities, allowing for standardized…
Custom Testing Scripts
AI Red Teamers frequently write custom scripts (often in Python) to automate bespoke attacks, interact with specific AI APIs, generate complex prompt sequences, parse model outputs at scale, or implement novel exploit techniques not found in standard tools. Proficiency in scripting is essential for advanced AI red…
Reporting Tools
AI Red Teamers use reporting techniques and potentially tools to clearly document their findings, including discovered vulnerabilities, successful exploit steps (e.g., effective prompts), assessed impact, and actionable recommendations tailored to AI systems. Good reporting translates technical findings into…
Lab Environments
AI Red Teamers need environments to practice attacking vulnerable systems safely. While traditional labs (HTB, THM, VulnHub) build general pentesting skills, platforms are emerging with labs specifically focused on AI/LLM vulnerabilities, prompt injection, or adversarial ML challenges.
CTF Challenges
Capture The Flag competitions increasingly include AI/ML security challenges. Participating in CTFs (tracked on CTFtime) or platforms like picoCTF helps AI Red Teamers hone skills in reverse engineering, web exploitation, and cryptography applied to AI systems, including specialized AI safety CTFs.
Red Team Simulations
Participating in or conducting structured red team simulations against AI systems (or components) provides the most realistic practice. This involves applying methodologies, TTPs (Tactics, Techniques, and Procedures), reconnaissance, exploitation, and reporting within a defined scope and objective, specifically…
LLM Security Testing
The core application area for many AI Red Teamers today involves specifically testing Large Language Models for vulnerabilities like prompt injection, jailbreaking, harmful content generation, bias, and data privacy issues using specialized prompts and evaluation frameworks.
Agentic AI Security
As AI agents capable of autonomous action become more common, AI Red Teamers must test their unique security implications. This involves assessing risks related to goal hijacking, unintended actions through tool use, exploitation of planning mechanisms, and ensuring agents operate safely within their designated…
Responsible Disclosure
A critical practice for AI Red Teamers is responsible disclosure: privately reporting discovered AI vulnerabilities (e.g., a successful jailbreak, data leak method, or severe bias) to the model developers or system owners, allowing them time to remediate before any public discussion, thus preventing malicious…
Specialized Courses
Targeted training is crucial for mastering AI Red Teaming. Look for courses covering adversarial ML, prompt hacking, LLM security, ethical hacking for AI, and specific red teaming methodologies applied to AI systems offered by platforms like Learn Prompting, Coursera, or security training providers.
Industry Credentials
Beyond formal certifications, recognition in the AI Red Teaming field comes from practical achievements like finding significant vulnerabilities (responsible disclosure), winning AI-focused CTFs or hackathons (like HackAPrompt), contributing to AI security research, or building open-source testing tools.
Emerging Threats
AI Red Teamers must stay informed about potential future threats enabled by more advanced AI, such as highly autonomous attack agents, AI-generated malware that evades detection, sophisticated deepfakes for social engineering, or large-scale exploitation of interconnected AI systems. Anticipating these helps shape…
Advanced Techniques
The practice of AI Red Teaming itself will evolve. Future techniques may involve using AI adversaries to automatically discover complex vulnerabilities, developing more sophisticated methods for testing AI alignment and safety properties, simulating multi-agent system failures, and creating novel metrics for…
Research Opportunities
AI Red Teaming relies on ongoing research. Key areas needing further investigation include scalable methods for finding elusive vulnerabilities, understanding emergent behaviors in complex models, developing provable safety guarantees, creating better benchmarks for AI security, and exploring the socio-technical…
Industry Standards
As AI matures, AI Red Teamers will increasingly need to understand and test against emerging industry standards and regulations for AI safety, security, and risk management, such as the NIST AI RMF, ISO/IEC 42001, and sector-specific guidelines, ensuring AI systems meet compliance requirements.
Questions about the AI Red Teaming Roadmap
Is the AI Red Teaming Roadmap free?
Yes. All 64 lessons in this roadmap are free to read on Bibha AI Learning Labs, with no account and no payment. Signing in is optional: it lets you mark lessons as done and keep your progress, and it opens the AI tutor, which runs on your own API key.
How long does the AI Red Teaming Roadmap take?
There is no fixed duration. It has 64 lessons in 5 stages, and the time it takes depends on how many hours you give it each week. Nothing is locked, so you can skip what you already know, and your progress is saved once you sign in.
Where should I start the AI Red Teaming Roadmap?
Start with stage 1, Introduction, and work down in order. The stages are numbered in the order the roadmap suggests, but every lesson is open, so if you already know Introduction you can go straight to Prompt Hacking or anything after it. Each lesson links to free resources for going deeper.
Can I learn AI Red Teaming in Hindi?
The lessons are written in English. Once you sign in, you can ask the AI tutor on any lesson to explain it in Hindi or Hinglish. It answers from that lesson’s own text, so the explanation stays on the syllabus. The tutor runs on your own Google, OpenAI or Anthropic key.