An Intuitive Guide on Agentic AI Security Threats -Part 2

Everything to Know about Key Security Threats in AI Agents

AI agents are getting adopted across organizations at a rapid pace. Marc Benioff of Salesforce mentioned in a recent interview that AI agents at Salesforce are now performing anywhere from 30% to 50% of the work. Satya Nadella mentioned in April 2025 that as much as 30% of code at Microsoft is AI-generated. In the same interview, Mark Zuckerberg stated that by 2026, half of the development will be done by AI, rather than people.

As the proliferation of AI agents dominates our lives and Enterprises, it is crucial to use them safely. To do so, we need to first understand the security threats related to AI Agents.

You are only as secure as your threat awareness, security awareness isn’t optional — it’s foundational to our survival and AI is no exception.

The attack surface for AI Agents comes from

• Input Prompts
• Large Language Models(LLMs)
• Tools or APIs
• Interactions with External entities

In Part 1 you became aware of the threats to AI agents via Prompts and LLM which includes Prompt Injections Attack, JailBreaking, Indirect Prompt Injection Attack, Backdoor Attacks, Misalignment, Hallucinations and Planning Threats