New research highlights a critical vulnerability in today's leading AI models. Cisco's head of AI threat intelligence, Amy Chang, revealed that multi-turn attacks, where attackers adapt their strategy over a conversation, successfully compromised AI models in up to 88.3% of cases. This starkly contrasts with traditional single-turn testing, which significantly underestimates the real-world risks.
The findings, presented at VB Transform 2026, underscore a growing concern among enterprises. A recent survey indicated that over half of businesses have experienced an AI security incident or a near-miss. Current security measures, such as scoped identities and sandboxing, are not universally adopted, lea ving many organisations exposed. This lack of robust security infrastructure is driving significant investment in AI security solutions, with major players like Palo Alto Networks, CrowdStrike, and Cisco making substantial acquisitions in the identity and isolation space.
Chang emphasised that single-turn testing, which involves a single malicious prompt, fails to capture the dynamic nature of real-world AI interactions. Multi-turn attacks, however, simulate how users actually engage with AI agents and applications, revealing vulnerabilities that a simple prompt would miss. The study tested 15 flagship AI models, with multi-turn success rates varying but consistently showing a non-trivial exposure across all tested models. Importantly, the ranking of model vulnerability differed significantly between single-turn and multi-turn testing methodologies.
The solution, according to Chang, lies in a return to fundamental security principles. Cisco's Integrated AI Security and S afety Framework provides a comprehensive approach to identifying and mitigating AI compromise points throughout the AI lifecycle. Experts from other companies, like Box's CISO Heather Ceylan, echo this sentiment, advocating for continuous multi-turn testing and pressure-testing agents to ensure execution controls function as intended. The key takeaway for enterprises is clear: to effectively secure AI, security testing must evolve to mirror the sophisticated, multi-turn attacks that adversaries employ.
Fuente Original: https://venturebeat.com/security/openai-anthropic-google-and-xai-models-all-broke-under-multi-turn-attack-up-to-88-of-the-time
Artículos relacionados de LaRebelión:
- Microsofts AI Models Slash Costs 89 Versus OpenAI
- AI Models Escaped OpenAIs Security Scare
- OpenAI Models Escaped Containment and Attacked Hugging Face
- China May Restrict Global Access To Top AI Models
- Critical NGINX Flaws Allow Remote Code Execution
Artículo generado mediante LaRebelionBOT
No hay comentarios:
Publicar un comentario