Meta has introduced CyberSecEval 3, a framework to assess the cybersecurity risks and capabilities of AI models. This comes in response to malicious actors increasingly targeting large language models (LLMs) and turning them into tools for cyberattacks. CyberSecEval 3 evaluates eight different risks across two categories: risk to third parties and risk to application developers and end users.
The latest version includes new areas focused on offensive security capabilities, such as automated social engineering and autonomous offensive operations. The CyberSecEval 3 team tested the LLM Llama 3 and found vulnerabilities in automated phishing and offensive operations. To combat these risks, Meta recommends several strategies.
First, deploy advanced guardrails like LlamaGuard 3 and PromptGuard. These tools have proven effective in reducing malicious code generation and the success rates of prompt injection attacks. Second, enhance human oversight.
Mitigating AI’s cybersecurity threats
The findings show that models still require significant human monitoring, especially in critical environments. Third, strengthen phishing defenses through AI detection tools and real-time monitoring to neutralize automated spear-phishing campaigns.
Fourth, invest in continuous training for cybersecurity teams to effectively use LLMs for defensive purposes and as part of red-teaming exercises. Finally, adopt a multi-layered security approach. Combining AI-driven insights with traditional security measures can significantly enhance an organization’s defense.
The CyberSecEval 3 framework provides actionable strategies for CISOs and cybersecurity leaders to mitigate the risks of weaponized LLMs. For organizations using LLMs in production, adopting these strategies can bolster their defenses against AI-driven cyberattacks. Staying proactive and informed is crucial to safeguard organizations against the advancing threats posed by weaponized LLMs.
Meta’s CyberSecEval 3 offers valuable insights and recommendations to help navigate this evolving landscape.
