Building safe AI agents
Generative AI, such as ChatGPT and Copilot, already helps us with daily tasks such as writing, summarising and finding information. But what happens when an AI agent is also allowed to perform actions on someone's behalf?
Generative AI-applications produce an answer or a recommendation. Agentic AI goes further – it can receive a goal, determine which steps are required, retrieve information, choose and use tools, interact with other systems, and perform actions.
That shift, from answering to acting, introduces a new set of risks that developers and product owners need to design for from day one.
This course is built for the people who design, build and ship agentic systems. Based on the OWASP Top10 for Agentic Applications – the industry framework for agentic AI risk – it walks the target audience through seven important questions that need answering before an AI agent is given the ability to act, and what safeguards need to be in place before it reaches production.
What are we building?
A shared understanding of what makes a system"agentic" – and how OWASP maps the risk landscape that comes with it.
What data and context can it retain?
Memory poisoning and context manipulation, and how to design agents that don't carry forward corrupted state.
How autonomous should it be?
Apply OWASP's "least agency" principle: autonomy should be earned deliberately, not granted by default.
When must a human take over?
Checkpoints that stop a single bad decision from cascading across connected agents and tools.
What can influence it?
How agents can be manipulated through prompts, poisoned data, or compromised tools – risks OWASP categorises under prompt injection and tool misuse.
Is it safe enough for production?
A practical, OWASP-aligned checklist for assessing whether an agentic AI system is ready to go live.
What can it access and do?
Design permissions and identity controls that prevent the privilege abuse OWASP flags as one of the top risks in agentic systems.
Explore the rest of our courses
Course in protection against AI manipulation
Our course on AI manipulation consists of short, 2–3-minute lessons on how we experience AI – both when we take the initiative ourselves and in situations where we may not even be aware of it. The course focuses on how AI technology is used by cybercriminals and the importance of keeping AI in mind in our daily lives.
Secure coding with OWASP Top 10
This course from Junglemap helps developers understand and reduce the most critical security risks in web applications. Applications are complex, and many vulnerabilities arise from everyday development decisions. Cybercriminals do not focus on how an application is intended to work, but on how it can be misused. This course helps you recognise these risks and understand how to prevent them in practice.
Information security awareness course for all employees
Our information security awareness training consists of 3-minute lessons on cyber security risks and how to counteract them, with topics such as social engineering, password and email safety, AI-manipulation, and how to avoid and handle ransomware attacks.
Privacy & GDPR course for all employees
Our Privacy awareness training covers what all employees need to know about privacy and privacy legislations such as the GDPR. The course consists of 10 content lessons, surveys and repetition quizzes, all designed to build and reinforce knowledge about privacy all year around.
