Securing AI and LLM features: prompt injection and beyond
Many applications now include AI chatbots, assistants and summarisers built on large language models (LLMs). These features bring new risks: prompt injection, data exfiltration and abuse that traditional security testing does not cover.
How it happens
An attacker writes input that overrides the model's instructions (direct prompt injection), or hides instructions in content the model reads, such as a web page or document (indirect prompt injection). The model may then reveal data, take unintended actions or produce harmful output.
Why it matters
- Leaks of system prompts, internal data or other users' information.
- Unauthorised actions when the model can call tools or APIs.
- High costs from abusive or looping requests.
How to prevent it
- Treat model output as untrusted; never execute it or render it without encoding.
- Give the model the minimum data and tool permissions it needs.
- Separate system instructions from user content, and filter both inputs and outputs.
- Rate-limit AI endpoints and cap request size.
How Vooki Pro helps
- Tests AI features for direct and indirect prompt injection.
- Checks for data exfiltration, model denial of service and training-data poisoning risks.
- Scans the web application and APIs around the AI feature in the same run.
- Its own AI analysis, ThreatLens, runs entirely on your machine, so your scan data is never sent to a third-party AI service.
Frequently asked questions
What is indirect prompt injection?
Instructions hidden in content the model reads, such as a web page or document, that make it act against the user or the application.
Can prompt injection be fully prevented?
Not reliably today. Reduce the impact by limiting the model's permissions and data, and treating its output as untrusted.
Try Vooki Pro free for 7 days
If your product has an AI feature, add it to your security testing. The trial unlocks every feature for 7 days, and scanning runs on your own machine.