Skip to content
Olivier Reuland

Adding AI to your SaaS - Security Risks and Opportunities

Using a new AI solution is no different to using any other 3rd party solution, with a few additional and important considerations.

SaaS products are bolting on image generation, chat, classification, and "AI assistants" faster than risk registers update. Useful for customers. Ugly when the vendor assessment still looks like a 2019 questionnaire.

Treat a new AI feature like any other third party first: data handling, security controls, financial health, compliance, continuity, incident response. If you do not have that process, start there - AI will not invent it for you.

Then add the failure modes the old questionnaire missed.

Extra risks that bite

  • Prompt injection - free-form input that tricks the model into leaking data, changing prices, or calling tools it should not
  • Sensitive output - DLP for what the model says, not only what users upload
  • Privacy and training - PII in RAG stores, prompts retained, data used to train the next model
  • Cross-tenant leak - your content showing up in someone else's answers
  • Regulated decisions - fairness, explainability, and lawful process when the model influences outcomes
  • Copyright - training data provenance and who gets sued when the output looks familiar
  • Reputation - wrong or insulting answers at customer scale
  • Unbounded consumption - attacker burns your API bill ("DDoS your wallet")

That last one is under-discussed. Classical injection assumes structured input. An LLM that accepts anything is a harder surface.

I walk the OWASP list in Securing Your LLM Applications with the OWASP Top 10. For provider vs weights risk, see Compare LLM Model vs LLM Service. For how to fold AI into ordinary risk work, Securing AI Initiatives.

Primary references worth keeping open:

Ship the feature when you can answer the questionnaire and the AI-specific questions - not when the demo looks clever.

Related

Olivier Reuland
Want to work with me? Check out WHC Services