Cisco's experiment reveals 88.3% multi-turn attack success rate across 15 models

According to findings presented by Amy Chang, Cisco's head of AI threat intelligence and security research, the company conducted a comprehensive red-teaming exercise involving 6,986 multi-turn attacks against 15 flagship AI models. The results were stark: attackers who adapted their prompts across the conversation broke through defenses as often as 88.3% of the time. This stands in stark contrast to the less than 20% success rate typically observed in single-turn attacks on the same models. The experiment, detailed during the agentic security panel at VB Transform 2026, highlights a fundamental flaw in current security testing practices. Many organizations still rely on single-turn evaluations, which Chang argued create a false sense of security. The sheer scale of the test — nearly 7,000 attack chains — provides robust statistical evidence that multi-turn vulnerabilities are widespread.

Why multi-turn attacks circumvent existing safeguards

Single-turn red-teaming evaluates a model's response to a single malicious prompt in isolation. However, real-world attackers — especially those using advanced AI agents — rarely operate that way. Instead, they build rapport, probe boundaries, and gradually steer the conversation toward prohibited outputs. Cisco's research demonstrates that models which initially refuse harmful requests can be tricked after several turns of seemingly benign dialogue. For example, an attacker might first ask about general safety procedures, then inquire about hypothetical exploits, and finally request specific code or instructions. Each individual request might pass a single-turn filter, but the cumulative effect bypasses the model's safeguards. This technique, sometimes called 'jailbreaking by attrition,' leverages the model's tendency to maintain conversational coherence. The 88.3% success rate indicates that nearly all tested models have insufficient multi-turn guardrails, posing a severe risk for enterprise deployments where AI agents interact with users over extended sessions.

Enterprise AI security survey reveals widespread concern but slow adoption

VentureBeat's June 2026 Pulse survey, which polled 107 enterprise respondents, provides context for why Cisco's findings resonated strongly with the audience at VB Transform 2026. According to the survey, 72% of enterprises using AI agents reported experiencing at least one security incident in the past 12 months. However, only 34% said they conduct regular multi-turn red-teaming exercises. A majority still rely on single-turn evaluations, often performed by third-party vendors. The survey also found that 61% of respondents considered 'conversational attack vectors' as a top-three security concern for the next year. Yet budget allocation tells a different story: only 28% of AI security budgets are directed toward adversarial testing. This gap between awareness and action is precisely what Amy Chang highlighted in her talk. She argued that until multi-turn testing becomes standard practice, enterprises remain exposed to sophisticated attackers who exploit conversational dynamics.

Implications for AI model deployment and regulation

Cisco's research arrives at a time when regulators worldwide are grappling with how to ensure AI safety. The European Union's AI Act, which entered full effect in 2025, requires rigorous testing for high-risk systems, but it does not explicitly mandate multi-turn adversarial evaluation. Similarly, the U.S. Executive Order on AI Safety from 2023 has been updated in 2025 to include 'conversational red-teaming' guidance, but enforcement remains voluntary. Chang suggested during the panel that CISOs and security teams should advocate for industry standards that incorporate multi-turn attack scenarios. She also noted that model providers — including OpenAI, Google, and Anthropic — have begun publishing system cards and red-teaming results, but most still focus on single-turn performance. Without a standardized multi-turn benchmark, enterprises cannot effectively compare the security posture of different models. The 88.3% failure rate across all tested models suggests that no provider currently offers adequate protection against adaptive conversational attacks, making this a systemic industry issue rather than a vendor-specific problem.

Recommendations for strengthening AI security programs

During the VB Transform 2026 panel, Amy Chang offered several actionable recommendations for enterprises. First, she urged security teams to implement multi-turn red-teaming as a core component of their AI evaluation pipeline, not as an optional add-on. Second, she called for the development of internal attack taxonomies that classify conversational strategies such as 'topic drifting,' 'context poisoning,' and 'gradual escalation.' Third, Chang recommended that organizations simulate attacker persistence — i.e., testing how many turns an adversary is willing to engage before succeeding. Her team at Cisco has open-sourced a limited version of their multi-turn testing framework, which includes 500 attack chains covering common patterns. The response from the community, she reported, has been overwhelmingly positive, with over 2,000 downloads in the first week. However, she cautioned that open-source tools are only a starting point; enterprises need dedicated red teams that understand both AI model behavior and adversarial psychology. The ultimate goal, she said, should be to reduce multi-turn attack success rates below 10%, a target that will require concerted industry effort.