A few years ago, we worried about what an AI system might say. Today, we must worry about what it might do.
Now imagine that colleague:
- Misunderstands its goal
- Trusts poisoned information
- Misuses a powerful tool
- Communicates with a compromised agent
- Executes code you never intended
- Continues acting without asking for approval
That is where agentic AI security becomes very real.
The OWASP Top 10 for Agentic Applications for 2026 helps us understand the risks that appear when AI systems can plan, remember, use tools, communicate and act on our behalf.
To make these risks easier to understand, I have created 10 short videos one for each risk:
- Agent Goal Hijack
- Tool Misuse and Exploitation
- Identity and Privilege Abuse
- Agentic Supply Chain Vulnerabilities
- Unexpected Code Execution
- Memory and Context Poisoning
- Insecure Inter-Agent Communication
- Cascading Failures
- Human-Agent Trust Exploitation
- Rogue Agents
These are not only “AI problems.” They are identity problems, access-control problems, supply-chain problems, human-trust problems and sometimes, very expensive business problems.
Security documents are useful, but sometimes you need someone to sit beside you and say, “Let us take this one risk at a time. What does it actually mean? How could it affect us? What should we do about it?” That is exactly why I created this short video series.
Think of the playlist as a guided journey through the life of an AI agent from the moment it receives a goal to the moment it uses tools, accesses identities, remembers information, communicates with other agents and acts on our behalf.
1. ASI01 – Agent Goal Hijack
The story begins with the agent’s goal. What if an attacker quietly changes that goal? The agent may still behave confidently and efficiently but now it is efficiently working for the wrong objective.
2. ASI02 – Tool Misuse and Exploitation
Tools are the agent’s hands. When an agent is given powerful tools without sufficient boundaries, a harmless instruction can lead to data exposure, unauthorised changes or destructive actions.
3. ASI03 – Identity and Privilege Abuse
Whose identity is the agent using when it acts? This video explores what happens when agents receive excessive permissions, inherit powerful credentials or operate without clear accountability.
4. ASI04 – Agentic Supply Chain Vulnerabilities
An agent rarely works alone. It depends on models, plugins, tools, frameworks, data sources and third-party services. One compromised dependency can quietly influence the entire agentic workflow.
5. ASI05 – Unexpected Code Execution
Sometimes an agent does not merely recommend code it writes, modifies or executes it. If untrusted input reaches that execution path, the impact can move rapidly from an AI problem to a full system compromise.
6. ASI06 – Memory and Context Poisoning
Memory helps an agent learn and remain useful. Unfortunately, poisoned memory can also teach it the wrong facts, preferences or instructions and those manipulations may survive long after the original attack.
7. ASI07 – Insecure Inter-Agent Communication
In multi-agent systems, one agent may trust another agent’s message without independently verifying it. A compromised agent can therefore spread false instructions or malicious data across the entire system.
8. ASI08 – Cascading Failures
Agentic systems are connected. One incorrect decision can trigger another tool, another agent and another automated action. A small failure at the beginning of the chain can become a serious incident by the end.
9. ASI09 – Human-Agent Trust Exploitation
People often trust confident, polished AI responses. Attackers can exploit that trust by manipulating agents into producing persuasive but dangerous recommendations, requests or approvals.
10. ASI10 – Rogue Agents
Finally, we meet the agent that no longer behaves as expected. Whether caused by compromise, misconfiguration, conflicting goals or inadequate controls, a rogue agent may continue acting until someone or something stops it.
Why should you watch this series?
You do not need to be an AI researcher to care about agentic security. If you are a developer, these risks can help you design safer agents. If you work in security, they can guide threat modelling, testing and monitoring. If you are a leader, they can help you ask better questions before approving an agentic AI deployment.
Most importantly, the OWASP Agentic Top 10 reminds us that securing an agent is not only about protecting the model. We must protect the complete journey:
the goal it receives, the context it trusts, the identity it represents, the tools it uses, the agents it communicates with and the actions it takes.
My suggestion is simple: Do not treat the Top 10 as a compliance checklist to be completed and forgotten. Take one risk, watch its short video and discuss it with your team:
- Could this happen in our environment?
- What would the impact be?
- Would we notice it?
- Could we stop it before the action was completed?
🎥 Watch the complete playlist: https://www.youtube.com/playlist?list=PLB_BXi-iL-Bk
📖 Explore the official OWASP framework: https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/ and bring your developers, security teams and business leaders into the conversation.
#OWASP #AgenticAI #AISecurity #GenAI #CyberSecurity #ArtificialIntelligence #SecureAI #AppSec #AIagents #ResponsibleAI OWASP® Foundation OWASP GenAI Security Project