“AI” is present in every sphere of our life now. So, why not the concepts of “AI” and “Security” working together? In that context, there can be two possibilities – “AI Security” and “Security for AI”. Both of them have different meanings. Let us see their differences now.
What is “AI Security”?
“AI Security” can also be worded as “AI for Security” which is making use of AI to enhance security. This involves using AI and its tools at every phase of security process. This increases productivity and helps InfoSec professionals do things in a much faster and less stressful way.
Here are some ways in which AI tools simplify the cyber security posture of any organization:
- AI powered ‘SOC’ processes which reduce alert fatigue
- Threat detection(can detect threats faster)
- AI tools enhance incident response
- Enhanced physical security
- Anamoly detection (to reduce fraud)
- Detecting phishing and malicious emails through AI tools
- Vulnerability management
- Real time monitoring ensures that breaches and attacks are caught much earlier
These are some ways in which “AI for security” works. Let us next see “What is security for AI”:
Security for AI:
‘Security for AI’ on the other hand is securing the AI system itself. It is quite an irony that the AI that is securing AI processes has to also be secured from attacks!
As the use of generative AI tools (like ChatGPT, Claude) have expanded and most of it us use it all the time, attacks against them have also increased with time. Security for AI involves securing the data that is used to train the models, the data pipelines through which the data is carried and the environment as well. These are some of the attacks that occur against LLMs from OWASP Top 10 LLM risks for the year 2025:
- Prompt injection attacks
- Sensitive information disclosure
- Supply Chain
- Data and model poisoning (‘poison’ the data that is used to train the AI models)
- Improper output handling
- Excessive agency
- System prompt leakage
- Vector and embedding weaknesses
- Misinformation
- Unbounded consumption
Reference: LLMRisks Archive – OWASP Gen AI Security Project
The best way to prevent and recover from attacks against LLMs is to
- a list of LLMs (like ChatGPT, Claude) that are used in a company
- have AI usage policies regarding the data that should be uploaded to an AI model
- monitor AI data that is generated
- Keep track of who accesses the AI models
These are some practices that help to keep LLMs safe from different types of attacks.
This post is for Blogchatter Half Marathon 2.0
That’s really informative. Thanks for sharing! 😊