Artificial intelligence (AI) is transforming enterprise security, but it is also creating new opportunities for attackers.
From prompt injection and data poisoning to model abuse and AI-assisted phishing, organizations are facing an expanding attack surface that traditional security controls were never designed to address.
To help security professionals better understand these risks, Max Robertson, Senior Content Engineer and Head of the AI Security Squad at TryHackMe, joined the eSecurity Planet Podcast for a hands-on demonstration of direct and indirect prompt injection attacks using TryHackMe’s interactive training platform.
In the video, Max walks through how these attacks work in practice and why AI security requires a different way of thinking than traditional application security.
Prompt injection remains one of the biggest AI security risks
Prompt injection has become one of the most discussed threats against large language models (LLMs), and for good reason.
As Robertson explains, the attack shares important similarities with SQL injection. In both cases, attackers attempt to mix trusted instructions with untrusted input, causing an application to execute behavior the developer never intended.
During the demo, Robertson shows a direct prompt injection attack against a simulated AI-powered automotive sales assistant.
By carefully crafting prompts that override the chatbot’s existing instructions, he demonstrates how an attacker can manipulate the AI into accepting a fictional $1 vehicle purchase.
While intentionally simplified for training purposes, the exercise illustrates how insecure AI assistants can be persuaded to ignore their original constraints.
Indirect prompt injection raises even greater concerns
The second demonstration highlights why some security professionals consider indirect prompt injection an even more significant enterprise risk.
Rather than attacking the chatbot directly, Robertson plants malicious instructions inside an external data source — in this case, a calendar entry that the AI assistant is authorized to access.
When another user later asks the assistant to summarize upcoming meetings, the malicious instructions are unknowingly ingested as trusted context, causing the AI to disclose confidential sales information.
Unlike direct prompt injection, the attacker does not need to be present when the exploit executes.
Any employee interacting with the AI assistant could unknowingly trigger the attack, making indirect prompt injection especially dangerous for AI agents connected to enterprise data sources.
AI attacks extend far beyond prompt injection
Although prompt injection dominates many AI security discussions, it represents only one category of emerging threats.
Organizations are also facing risks such as:
- Data poisoning, where attackers manipulate training or retrieval data to influence model outputs.
- Model inversion and extraction, which attempt to recover sensitive training data or replicate proprietary AI models.
- Jailbreaking, where carefully crafted prompts bypass model safety controls.
- AI-assisted phishing and social engineering, which use generative AI to create increasingly convincing, personalized attacks.
- Malicious AI agents, capable of autonomously interacting with enterprise applications, APIs, and cloud services.
As enterprises increasingly connect AI systems to sensitive business data and operational workflows, weaknesses in these integrations can have consequences that extend well beyond inaccurate chatbot responses.
AI security starts with understanding the technology
Robertson emphasizes that AI should not be viewed as something to fear, but rather something to understand.
Just as organizations adapted to cloud computing and other major technology shifts in the past, defenders now need to build practical knowledge of AI systems, their capabilities, and their security implications.
TryHackMe’s hands-on labs are designed to help practitioners safely explore realistic AI attack scenarios before encountering them in production.
Readers interested in developing hands-on AI security skills can also save 30% on TryHackMe with coupon code TECHADVICE30.
As AI adoption accelerates, organizations need strong governance, secure development, and continuous training to defend against evolving AI threats.
Prompt injection is just one of many risks security teams must prepare for.
Extending Zero Trust principles to AI systems can help limit the impact of compromised prompts, identities, and data sources.