Securing AI Against Evolving Threats

This report surveys AI-specific cyber risks—such as data poisoning, backdoor attacks, inference/extraction attacks, and prompt injection—reviews security benchmarking (e.g., CyberSecEval2) and frameworks (MITRE ATT&CK/Atlas, OWASP LLM, NIST AI RMF, Google SAIF), and recommends adopting benchmarks and security-driven frameworks to assess and mitigate AI model risks while describing Arctic Wolf’s approach to using these tools to harden AI deployments.