Traditional security tools were built for predictable systems and known attack patterns. GenAI environments behave very differently and introduce new, AI-native risks that legacy tools struggle to see or interpret.

Several factors drive this gap:

• Unpredictable outputs: GenAI models generate highly variable text, code, and images that don’t follow fixed signatures. Static rules and pattern matching miss threats hidden in natural language.
• New attack types: GenAI brings risks such as prompt injections, data poisoning, model tampering, and agent hijacking that don’t exist in traditional IT stacks.
• Limited visibility: Conventional tools rarely monitor model inputs, outputs, and behavior in real time, which makes it hard to detect subtle misuse or unsafe responses.
• Unstructured data: GenAI frequently processes text, code, and images—formats that many legacy tools don’t analyze deeply enough for AI-specific threats.

The impact is already visible. 57% of organizations report a rise in AI-driven attacks, and one large cloud provider saw daily cyber threats jump from 100 million to nearly 1 billion in six months, partly due to AI. To keep pace, security teams need AI-native controls that understand models, prompts, agents, and data—not just networks and endpoints.

GenAI reshapes how enterprises work, but it also introduces a distinct set of risks that go beyond traditional vulnerabilities. Key issues include:

• Prompt injection: Attackers craft prompts that tell the model to ignore its guardrails (for example, “ignore the restrictions on sharing customer information”). Because this looks like normal usage, it’s hard to detect and can lead to data leaks, policy violations, and “jailbroken” behavior.
• Open-source model risks: Public models can be altered or packaged with malicious code. Deserialization attacks (e.g., via pickle or joblib) can trigger remote code execution, while model tampering can embed backdoors or hidden triggers that only appear under certain conditions.
• Insecure outputs: Models may generate harmful URLs or content, sometimes based on poisoned or compromised sources. A single malicious link can open the door to phishing, malware, or unauthorized access—especially when users trust AI-generated suggestions.
• Sensitive data leaks: GenAI systems often operate on proprietary data that sits outside traditional controls. If attackers breach the application or manipulate the model, they can access intellectual property or confidential information.
• Hallucinations and false citations: Models can confidently invent research, sources, or facts. In professional or regulated settings, these fabricated citations can mislead decisions and undermine trust.
• Agent hijacking: With 93% of IT leaders planning to introduce autonomous AI agents and 78% of companies planning to use them broadly, agents are becoming high-value targets. Indirect prompt injections can hide malicious instructions in data the agent reads, causing it to take harmful actions (for example, exfiltrating emails) while appearing to behave normally.

Because these threats are embedded in prompts, training data, model files, and agent behavior, they require security that is purpose-built for GenAI—not just extensions of existing tools.

Prisma AIRS is designed to help organizations secure GenAI in a more integrated, AI-native way, rather than relying on disconnected point tools. It focuses on protecting models, data, applications, and agents across the full lifecycle.

Core capabilities include:

• AI Model Scanning: Scans models—especially from open-source and managed services—to detect malicious code, backdoors, insecure configurations, and deserialization risks before deployment. This helps ensure models are safe, trustworthy, and aligned with your security policies.
• Posture Management: Provides continuous visibility into how AI systems and agents are configured and used. It helps you:
  • Identify and remediate misconfigurations
  • Enforce least-privilege access for agents
  • Guard against non-compliant or unsafe usage
• AI Red Teaming: Supports structured testing of your AI ecosystem, surfacing vulnerabilities and providing real-time recommendations to strengthen defenses.
• AI Agent Security: Applies guardrails to prompts and responses to block harmful or toxic content, set topic limits, and reduce hallucinations, helping agents operate safely even as they act autonomously.
• Runtime Security: Monitors AI applications and models in production to prevent malware from entering, block execution of malicious code, and protect AI workloads at runtime.

Instead of stitching together multiple point products, Prisma AIRS brings these capabilities into a single platform. This integrated approach helps security teams keep up with fast-moving GenAI adoption while maintaining control over risk, compliance, and overall security posture.