Inside Cisco's 2025 AI Security Report: Emerging Threats and Defense Strategies

Rapid AI Adoption Outpaces Security Preparedness

As AI technologies become integral to business operations, security challenges are intensifying. Cisco's "State of AI Security in 2025" report reveals that while 72% of organizations have adopted AI, only 13% feel fully prepared to secure it effectively. This gap underlines the urgent need to address evolving AI-specific threats beyond traditional cybersecurity approaches.

New and Evolving AI Threats

AI infrastructure faces increasing attacks, such as the compromise of NVIDIA's Container Toolkit and the Ray AI framework. These breaches demonstrate how vulnerabilities in AI components can lead to widespread risks. Additionally, supply chain attacks like "Sleepy Pickle" exploit open-source AI components, making tampering and detection difficult. Emerging attack methods include prompt injection, jailbreaking, and training data extraction, which allow attackers to bypass safety controls and access sensitive data.

Attack Vectors Exploiting AI Systems

Malicious actors target AI throughout its lifecycle—from data collection and training to deployment. Jailbreaking uses adversarial prompts to circumvent model safeguards, while indirect prompt injection manipulates input materials like malicious PDFs to provoke harmful outputs without direct system access. Training data extraction exposes proprietary information, and data poisoning, even at fractions of a percent of large datasets, can significantly alter AI behavior with minimal cost.

Key Findings from Cisco’s Research

Cisco's team demonstrated successful automated jailbreaking of leading models using Tree of Attacks with Pruning (TAP). Fine-tuned models are notably more vulnerable, being over three times easier to jailbreak and more prone to harmful content. Their research also shows how simple techniques can extract sensitive training data and how inexpensive data poisoning can meaningfully impact AI models.

AI as a Tool for Cybercrime

AI not only faces threats but also empowers cybercriminals. Automation and AI-driven social engineering increase the sophistication and effectiveness of attacks like phishing and voice cloning. Malicious tools such as "DarkGPT" enable even low-skilled attackers to craft personalized exploits that evade traditional defenses.

Recommendations for Strengthening AI Security

Cisco advises organizations to manage risks across the entire AI lifecycle, secure third-party components, and implement traditional cybersecurity best practices like access control and data loss prevention. Focusing defenses on vulnerable points like supply chains and educating employees on AI risks are also critical steps.

The Future of AI Security

As AI adoption accelerates, security risks will evolve in complexity. Governments and organizations are beginning to formulate policies to balance innovation with safety. Prioritizing security alongside AI development will be key to leveraging AI’s benefits while mitigating threats in the years ahead.