All 60 tasks, level by level

The real graded AI Red Teaming tasks the portal assigns, grouped into three levels.

The portal assigns these AI Red Teaming tasks. Finish Level 1 before Level 2 and Level 2 before Level 3. Your mentor may mark some tasks as required for the certificate.

Level 1: Foundations (24 tasks)

Task	What to do
What AI red teaming is	Explain AI red teaming, why it matters, and the rules of ethical testing: only test systems you own or are permitted to test.
How LLMs work	Explain tokens, prompts, system prompts, context windows and why models can be manipulated.
Threat landscape	Summarise the OWASP Top 10 for LLM Applications with one example per risk.
Set up a safe lab	Set up a local test app with a small open model (Ollama) or a sandboxed API key, in a repo named ai-redteam-internship.
Testing plan	Write a test plan: scope, target behaviours, success criteria and what is out of bounds.
Prompt injection basics	Build a small summariser app and show how instructions hidden in input text can change its behaviour.
System prompt leakage	Test whether your own app reveals its system prompt and record the prompts that worked.
Jailbreak categories	Classify jailbreak styles (role play, obfuscation, multi-turn) without publishing harmful outputs.
Hallucination testing	Write 20 factual questions with known answers and measure how often the model is wrong or makes things up.
Bias probing	Test the model with paired prompts that differ only in gender, caste or religion and compare the answers.
Logging findings	Create a findings template: ID, prompt, response, severity, reproducibility and fix.
Responsible disclosure	Explain responsible disclosure and how to report a model issue to a vendor's bug bounty or safety team.
AI incident database	Read 10 real AI incidents from a public incident database and classify the failure types.
Harm taxonomy	Build a taxonomy of AI harms (misinformation, privacy, bias, unsafe advice) with examples.
Content policy reading	Summarise the usage policy of two AI providers and list what is not allowed.
Prompt hygiene	Review five system prompts for secrets, weak instructions and missing rules.
Refusal quality	Test how a model refuses harmful requests and rate whether refusals are helpful and polite.
Over-refusal testing	Find safe requests that a model wrongly refuses and report them.
Factuality test set	Build a set of 30 questions about Indian facts with verified answers and test a model.
Citation checking	Check whether a model's cited sources actually exist and support its claims.
Personal data handling	Test whether a chatbot asks for or stores personal data unnecessarily.
Report writing practice	Rewrite three vague findings into clear, reproducible reports.
Safe lab hygiene	Document how you isolate lab models, keys and logs from personal accounts.
Threat actor profiles	Write profiles of likely misuse actors for an education chatbot.

Level 2: Building real features (20 tasks)

Task	What to do
Indirect prompt injection	Build a RAG demo where a planted document tries to hijack the answer, and show it working on your own app.
Defences against injection	Add input separation, output checks and least-privilege tools to your app and retest.
Tool and agent abuse	Build a small agent with one harmless tool and test whether injected text can make it misuse the tool.
Data leakage	Test whether your RAG app leaks documents a user should not see, then add access filtering.
Automated red teaming	Use an open-source tool (garak, PyRIT or promptfoo) to run an automated scan on your lab model.
Evaluation suite	Build a repeatable test suite of attack prompts with pass/fail checks and run it in CI.
Guardrails	Add a guardrail layer (moderation API, classifier or rules) and measure what it blocks and wrongly blocks.
Multilingual attacks	Test whether safety behaviour changes when prompts are in Tamil, Hindi or mixed languages.
Severity scoring	Define a severity scale for AI findings and score all your findings with it.
Findings report	Write a professional red team report on your lab app with an executive summary and fixes.
Encoding and obfuscation tests	Test whether filters can be bypassed with encodings or misspellings in your own lab app, and fix it.
RAG poisoning	Plant a misleading document in your RAG lab and measure how often it changes answers.
Agent tool permissions	Audit an agent's tools for dangerous capabilities and add confirmation steps.
Output handling	Test whether model output can inject HTML or scripts into your web UI, and sanitise it.
Rate and cost abuse	Test how an attacker could run up your API bill and add limits.
Moderation benchmark	Compare two moderation tools on a labelled set and report precision and recall.
Bias metrics	Measure response differences across groups with a simple metric and chart it.
Image model testing	Test an image generation or vision model for unsafe or biased outputs in a controlled way.
Red team playbook	Write a playbook with test categories, steps and severity rules for your team.
Fix verification	Retest five fixed findings and confirm they stay fixed after a model update.

Level 3: Professional work and capstone (16 tasks)

Task	What to do
Multi-turn attack testing	Design multi-turn conversations that slowly push the model off-policy and measure when defences fail.
Model supply chain	Review the risks of downloading third-party models and datasets (poisoning, unsafe pickles) and how to check them.
Adversarial examples	Create adversarial inputs for a small image or text classifier and measure the accuracy drop.
Privacy attacks concept	Explain membership inference and training-data extraction and how to reduce the risk.
Red team exercise	Run a timed red team exercise on a teammate's AI app (with permission) and swap reports.
Defence in depth design	Design layered defences for a production chatbot: input, model, output, tools, monitoring and human review.
Monitoring for abuse	Add logging and alerting that flags likely attacks in production traffic.
Capstone: AI security assessment	Deliver a full assessment of an AI app you built: threat model, automated and manual testing, fixes, retest results and a final report.
Automated adversarial generation	Use one model to generate test attacks against another in your lab and filter the results.
Long-context attacks	Test whether instructions buried deep in long documents affect your app.
Multi-agent risks	Test how one agent's output can mislead another in a two-agent system.
Evaluation in CI	Run your safety regression suite automatically on every prompt or model change.
Governance mapping	Map your findings to the NIST AI Risk Management Framework categories.
Responsible disclosure drill	Write a disclosure report for a finding as if submitting to a vendor's program.
Executive briefing	Brief a non-technical audience on AI risks of your app in 5 minutes.
Capstone 2: red team engagement	Run a full red team engagement on a teammate's AI product with rules of engagement, testing, report, fixes and retest results.

← Tools and first-week setup Submitting tasks and earning your certificate →