From the source
Perplexity published a guide to LLM security that walks through the OWASP Top 10 for LLM Applications 2026, from prompt injection to improper output handling.
- The company says it is skipping any claim that the problem is solved, citing OWASP project co-lead Wilson: "...frankly, I don't think we are at 99%."
- The guide's framing is that models will sometimes take orders from the wrong person, so defenses should sit around the model rather than inside it.
- The 2026 edition tells builders to "stop trying to build a model that cannot be fooled" and design the surrounding system so a fooled model cannot break anything important.
- Each of the ten risks gets a plain-English explanation, a real-world example and the practices that can prevent it or limit the damage.
- Perplexity points to its own experience with agents, quoting CISO Kyle Polley that agents "might not follow policies even if you add it to the system prompt," and to Numbat, its open-source tool that checks an agent's planned action and can block it when a configured security rule is broken.