Multi-turn attacks broke AI models 88% of the time
Cisco's research revealed the inadequacy of single-turn testing.
1Impact of multi-turn attacks
Cisco's research on 15 leading models found that multi-turn attacks achieved success rates as high as 88.3%. This highlights the inadequacy of single-turn tests in evaluating AI security.
2Corporate security gaps
According to a VentureBeat survey, 54% of companies experienced a security incident or narrowly avoided one. Yet only 32% implement comprehensive identity management for each agent, and 30% isolate high-risk agents.
3Major industry moves
Security giants like Palo Alto Networks, CrowdStrike, and Cisco made billion-dollar acquisitions to strengthen identity and isolation layers, reflecting the industry's efforts to address these gaps.
4Amy Chang's expertise
Amy Chang, Cisco's head of AI threat intelligence, led cybersecurity operations at JPMorgan Chase and served in the U.S. Navy. Her experience lends credibility to the findings.
5Research details
The study by Chang and Nicholas Conley tested 30,090 single-turn and 6,986 multi-turn attacks. Multi-turn success rates ranged from 7.89% to 88.3%, and model rankings differed between test types.
6LLM Security Leaderboard
Cisco publishes adversarial evaluation signals for 105 models, helping companies identify vulnerabilities and refine their security strategies.
7Importance of realistic testing
Chang emphasized that single-turn tests are insufficient, while multi-turn attacks better reflect real-world interactions, making it easier to detect harmful outputs and misaligned behaviors.
8Cisco's security framework
Cisco's Integrated AI Security and Safety Framework details how AI can be compromised throughout its lifecycle, enabling companies to develop strategies based on real incidents.
9Box's three-layer approach
Box CISO Heather Ceylan adopted a three-layer security approach: permissioning, ephemeral sandbox environments, and runtime control. This prevents unauthorized agent actions.
10Critical role of monitoring
Ceylan noted that a single mistake can erase all trust, making continuous monitoring vital. Models can change, and monitoring helps control the impact of these changes.
Original article: VentureBeat — Çoklu döngü saldırıları yapay zeka modellerini %88 oranında kırdı