Agentic Cybersecurity and Agentic Penetration Testing Course
Build, run and break AI security agents in live labs, and learn to defend the agents businesses are already deploying.
AI agents can now plan a task, call tools and act on the results. Security teams use them to speed up testing, and companies are deploying them in support desks, coding pipelines and internal workflows, often with access to data and systems that attackers want. Agentic Cybersecurity and Agentic Penetration Testing is a 12-week live course from Cehpoint School of Advanced Studies that teaches both sides of that shift.
On the offence side, you build and direct AI agents that carry out reconnaissance, vulnerability discovery, exploitation and reporting against authorised lab targets, with scope enforced in code and a human approval step before anything risky. You learn to build your own security agents, expose tools through Model Context Protocol style servers, and orchestrate planner, executor and verifier agents.
Then you turn around and attack AI systems themselves:
- Prompt injection, direct and indirect, through web pages, documents and tickets
- Agent goal hijacking, tool misuse and tool poisoning
- RAG and memory poisoning, and data exfiltration through agent actions
- Defending agents with least privilege, sandboxing, egress control, audit logs and kill switches
Every attack and defence is mapped to the OWASP Top 10 for LLM Applications 2025, the OWASP Top 10 for Agentic Applications for 2026 and MITRE ATLAS, so your findings read the way clients and auditors expect.
The course also covers what most training skips: measuring how accurate your agent's findings really are, controlling false positives, keeping model costs in check, and keeping a tamper-evident chain of custody for agent output so your evidence stands up to scrutiny.
All work happens in isolated labs or on systems you are authorised in writing to test, in line with India's IT Act, 2000. The course ends with a capstone engagement against a lab that combines a traditional application and an AI agent. You need to be 18 or older, with basic networking, Linux and scripting, or a completed ethical hacking course. The current fee is shown when you enrol.
What you will be able to do
- Design and run AI-agent workflows for reconnaissance, vulnerability discovery, exploitation and reporting against authorised lab targets
- Build security agents that call real tools through tool servers (including Model Context Protocol style servers) and orchestrate several agents with a human approval step
- Attack AI applications and agents in labs: direct and indirect prompt injection, agent goal hijacking, tool poisoning, RAG and memory poisoning, and data exfiltration through agents
- Map AI attacks and defences to the OWASP Top 10 for LLM Applications 2025, the OWASP Top 10 for Agentic Applications for 2026 and MITRE ATLAS
- Harden agents with least-privilege tool access, sandboxing, output validation, logging and human-in-the-loop controls
- Evaluate agent findings rigorously: reproduce them, control false positives, cap cost and keep a verifiable chain of custody for agent output
- Work within Indian law and written rules of engagement, and write a professional report that a client can act on
Syllabus, week by week
Foundations: how AI agents work, and the rules of engagement
- LLMs, agents, tools, memory and planning loops explained for security people
- Where agents help in offence and defence today, and where they fail
- Indian law for security testers: hacking a computer system is an offence under the IT Act, 2000; only test with written authorisation
- Scope documents, rules of engagement, and personal-data handling under the DPDP Act, 2023
- Setting up your isolated lab
Hands-on: Build your personal lab: an isolated network of intentionally vulnerable targets, a local agent runtime and logging, plus a signed scope template you will use for the rest of the course.
Agentic reconnaissance and vulnerability discovery
- Giving an agent tools: scanners, HTTP clients, DNS and subdomain tools
- Planning versus brute force: how to stop an agent wandering out of scope
- Scope enforcement in code, not just in the prompt
- Agent-assisted web and API testing against lab applications
- Triage: separating real issues from noise
Hands-on: Run a recon agent against the lab range with hard scope guards, then compare its findings with a manual baseline and record misses and false positives.
Agent-assisted exploitation and reporting in authorised labs
- Human approval gates before any state-changing action
- Exploit validation and safe proof-of-concept on lab targets
- Privilege escalation and lateral movement with an agent as co-pilot, not autopilot
- Evidence-backed reports: reproducible steps, screenshots, request logs
- Cost control: token budgets, caching, stopping runaway loops
Hands-on: Chain a multi-step compromise of a lab environment where the agent proposes each step and you approve it, then produce a client-grade report with full evidence.
Building and orchestrating security agents
- Tool servers and the Model Context Protocol: exposing tools safely
- Multi-agent patterns: planner, executor, verifier and reporter roles
- State, memory and handoffs between agents
- Defensive agents: log triage, alert enrichment and detection-rule drafting
- Testing your own agents before you trust them
Hands-on: Build a small multi-agent system with a planner, a tool-using executor and an independent verifier agent, exposed through your own tool server with per-tool permissions.
Attacking AI systems and agents
- OWASP LLM Top 10 2025 in practice: prompt injection, sensitive information disclosure, improper output handling, excessive agency, system prompt leakage, vector and embedding weaknesses, unbounded consumption
- Indirect prompt injection through web pages, documents, emails and tickets
- Agent goal hijacking, tool misuse and identity and privilege abuse
- Tool poisoning, malicious tool descriptions and agentic supply chain risk
- RAG and memory poisoning
- Data exfiltration through agents: links, images, tool calls and outbound requests
- Guardrail bypass techniques, studied only against lab systems you control
Hands-on: Red-team a deliberately vulnerable AI assistant and agent in the lab: achieve indirect prompt injection, poison its retrieval store, hijack a tool call and demonstrate a controlled exfiltration path, mapping each step to OWASP and MITRE ATLAS.
Defending agents
- Least privilege for tools and credentials; identity for agents
- Sandboxing, network egress control and allow-lists
- Input and output validation, content provenance and trust boundaries
- Monitoring agent actions: audit logs, anomaly signals and kill switches
- Threat modelling an agentic application with OWASP and MITRE ATLAS
- Incident response for agent misuse, and checking an organisation's CERT-In reporting obligations
Hands-on: Take the vulnerable agent from weeks 8-9, fix it, and re-run your own attacks to prove which defences hold and which do not.
What practitioners rarely teach
- Evaluating agent findings: ground-truth sets, precision and recall for your own tools
- False positive control and confidence scoring
- Reproducibility: versioned prompts, models and tool versions
- Chain of custody for agent output: hashing, timestamps and tamper-evident logs
- Running agents on a budget without cutting corners on evidence
- Ethics: what you must refuse to automate
Hands-on: Build an evaluation harness for your recon agent, measure its precision on a labelled lab dataset, and produce a hash-chained evidence log for one finding.
Capstone
- Scoping and rules of engagement for a full lab engagement
- Agentic assessment of a combined target: a web application plus an AI agent
- Executive summary and technical report
- Presentation and review
Hands-on: Capstone: a full authorised engagement against a lab that combines a traditional application and an AI agent, delivered with a scope document, evidence log, findings mapped to OWASP and MITRE ATLAS, remediation advice and a live defence of your work.
Tools you will use
Who this is for
Security analysts, penetration testers, SOC engineers, developers building AI features, and graduates or final-year students with a security foundation who want practical skill in using AI agents for offence and defence, and in testing and securing AI systems themselves.
Before you join
Must be 18 or older. Comfortable with basic networking (TCP/IP, ports, HTTP, DNS) and the Linux command line, with some scripting (Python preferred). Completing an ethical hacking course, such as the one at Cehpoint School of Advanced Studies, meets this. A laptop or desktop able to run virtual machines or containers, with a stable internet connection.
Safety and rules
Everything in this course is done only on systems you own or are explicitly authorised in writing to test, and on the isolated labs provided. Hacking a computer system is an offence under India's Information Technology Act, 2000. Learners agree to a code of conduct at enrolment, and using course techniques against real targets without authorisation leads to removal from the course. Guardrail-bypass and exfiltration techniques are taught so that defenders understand them, and are practised only in the lab.
Certificate
Learners who complete the labs and pass the capstone receive a Cehpoint School of Advanced Studies certificate of completion listing the skills assessed. It is a course certificate, not a government licence or an independent professional certification.
Frequently asked questions
What does agentic penetration testing mean?
It means using AI agents, software that plans, calls tools and acts in steps, to help with reconnaissance, testing, exploitation and reporting, while a qualified person sets the scope, approves risky actions and checks every finding. You also learn to test AI agents themselves as a target.
Will AI agents replace penetration testers?
Agents speed up repetitive work, but they also miss things, report findings that are not real and can wander out of scope. The skill that matters is directing agents, verifying their output and taking responsibility for the result, and that is what this course trains.
Is this legal?
Yes, when done properly. Every exercise runs on isolated labs or on systems you own. You learn to work only from written authorisation and rules of engagement, because hacking a computer system without it is an offence under the IT Act, 2000.
Which frameworks does the course follow?
Attacks and defences are mapped to the OWASP Top 10 for LLM Applications 2025, the OWASP Top 10 for Agentic Applications for 2026 and MITRE ATLAS, so your reports use the language clients and auditors recognise.
Do I need to know machine learning?
No. You need networking, Linux and basic scripting. The course explains how language models and agents work from a security engineer's point of view.
Do I need a paid AI subscription?
Lab access is arranged for the course. Where an exercise uses a hosted model, the course explains the options and how to keep usage and cost under control.
How much does the course cost?
The fee is shown on the enrolment page at the time you enrol. Fees are reviewed regularly, so check the current fee before booking.
Start with a free conversation
Opening an admission is free. An academic counsellor reads your goals and answers your questions; you pay only when you decide to enrol.