As artificial intelligence (AI) becomes a cornerstone of modern business practices, it simultaneously introduces a range of safety and security challenges that are evolving faster than traditional cybersecurity measures can keep up with.
The implications of these developments are profound, with potential risks that could have far-reaching consequences. A recent report from Cisco’s 2024 AI Readiness Index reveals that merely 29% of organizations surveyed feel adequately prepared to identify and thwart unauthorized manipulations involving AI technologies.
Ongoing Model Validation
DJ Sampath, the Head of AI Software & Platform at Cisco, emphasizes the importance of continuous model validation in this rapidly changing landscape. He states, “Model validation is not a one-off task; it requires ongoing attention.”
He elaborates on this point: “As we observe changes in the model—whether through fine-tuning or emerging threats—we must continuously gather insights and revalidate how these models respond to new types of attacks.”
Cisco’s advanced threat research team plays a crucial role in this process by actively monitoring AI-related attacks and exploring ways to enhance defenses against them. Sampath notes their contributions to standards organizations like MITRE, OWASP, and NIST as part of their commitment to improving security protocols.
In addition to mitigating harmful outputs from AI systems, Cisco is also focused on protecting these models from external threats that could alter their functionality. Risks such as prompt injection attacks, jailbreaking attempts, and training data poisoning necessitate robust preventive strategies.
The Complexity of Evolving Security Needs
Frank Dickson from IDC offers insights into how cybersecurity has transformed over time due to advancements in technology. He points out significant shifts: “Initially we transitioned from on-premise solutions to cloud-based systems which introduced an array of new challenges. As applications evolved from monolithic structures into microservices architectures, additional complexities emerged.”
He continues by highlighting the impact that artificial intelligence—and particularly large language models (LLMs)—has had on security concerns: “With each evolution comes an entirely new set of issues.”
The intricacies surrounding AI security increase as applications adopt multi-model frameworks where vulnerabilities can manifest at various levels—from individual models down through application layers—affecting developers, end-users, and vendors alike.
Dickson explains further: “Once an application migrates into the cloud environment—be it AWS or Azure—it typically remains there without frequent transitions between different platforms.” This stability contrasts sharply with earlier development practices where applications frequently shifted environments.
“Similarly,” he adds regarding microservices architecture,” once you establish your application within Kubernetes or another framework; you don’t revert back easily.”
When securing LLMs specifically, Dickson stresses the need for awareness about model changes—not just minor revisions but substantial shifts where developers might switch between entirely different frameworks like Anthropic or Gemini weekly. Each model presents unique strengths along with distinct vulnerabilities.
To address these multifaceted challenges effectively within diverse environments rather than relying solely on traditional safety measures tied to individual models,Cisco has launched its innovative AI Defense solution—a self-optimizing system utilizing proprietary machine learning algorithms designed for identifying evolving safety concerns informed by threat intelligence sourced from Cisco Talos.
Adapting To Rapid Advancements
Jeetu Patel serves as Executive VP and Chief Product Officer at Cisco; he shares his perspective regarding how swiftly technological advancements can feel revolutionary yet quickly become commonplace.
He draws parallels using examples such as Waymo’s self-driving cars: “Initially stepping into one feels surreal—there’s no driver! But after experiencing it multiple times? You start noticing things like seat comfort.”
Patel reflects upon society’s rapid acclimatization towards innovations like ChatGPT over recent years: “Any major breakthrough may seem extraordinary initially but soon becomes normalized.”
Looking ahead toward Artificial General Intelligence (AGI), Patel believes normalization will occur similarly but cautions against underestimating progress made by current models which unlock unprecedented use cases previously thought unattainable.
“No one anticipated smartphones would possess more computational power than mainframe computers right at our fingertips,” he remarks while observing how seamlessly integrated they’ve become into daily life—even among younger generations who take them for granted without much thought!
“It’s imperative for companies today not only recognize this shift but adapt swiftly alongside technological evolution.”
Student at Government College of Engineering Kolkata and an enthusiastic Journalist. Follows Topics such as Indian Politics, Gaming, Anime, etc. Click on the bell icon for more news and opinions from Srijan.