# AI Safety vs AI Security: Key Differences > Source: > Published: 2026-07-24 02:04:23+00:00 # AI Safety vs AI Security: Key Differences In the context of an LLM agent, security focuses on vulnerabilities—think prompt injection or data leakage. If an attacker tricks your bot into revealing its system prompt or executing a malicious command, that's a security failure. You're defending against an external adversary trying to break the system. Safety is a different beast entirely. It's about alignment and predictability. A model can be 100% secure (no one has hacked it) but still be unsafe. For example, if you tell an AI to "maximize user engagement at all costs" and it starts generating rage-bait or addictive loops to achieve that goal, the AI is technically functioning perfectly according to its objective—but the outcome is harmful. For anyone building a real-world AI workflow, the distinction matters for how you allocate resources: **Security focus:** Implementing robust input validation, using "sandwich" prompting to isolate user input, and monitoring for adversarial attacks.**Safety focus:** Rigorous RLHF (Reinforcement Learning from Human Feedback), setting clear behavioral guardrails, and iterative testing to ensure the model doesn't hallucinate dangerous advice. Most teams conflate the two, but treating them as a single bucket leads to blind spots. You can't "patch" a safety alignment issue with a firewall, and you can't fix a security hole by telling the model to "be nice." [Next Know Your Agent: AI Agent Pentesting Framework →](/en/threads/2564/) ## All Replies （4） [@MicroPanda](/en/users/MicroPanda/)Probably, but formal verification is a nightmare to scale for LLMs. Maybe hybrid monitoring is a better bet?