Photo credit: www.darkreading.com
Cisco is enhancing its cloud security solutions with a progressive technology aimed at enabling developers to identify and address vulnerabilities in artificial intelligence (AI) applications along with their foundational models.
The newly unveiled Cisco AI Defense, launched on January 15, focuses on mitigating data leaks, especially those involving employees using generative AI services such as ChatGPT, Anthropic, and Copilot. Currently, this offering is available to select early-access customers, with plans for broader availability set for March.
AI Defense functions within the Cisco Secure Access framework, part of the comprehensive secure service edge (SSE) cloud security suite introduced by Cisco last year. This software-as-a-service suite encompasses various features, including zero-trust network access, VPN-as-a-service, a secure web gateway, a cloud access security broker, firewall-as-a-service capabilities, and tools for digital experience monitoring.
Administrators can manage the AI Defense through the unified Cisco Cloud Control interface, which centralizes all of Cisco’s cloud security features.
Gaps in AI Capabilities
The purpose of AI Defense is to assist organizations grappling with the security challenges of integrating AI into their operations. Jeetu Patel, Cisco’s chief product officer and executive vice president, emphasized at the launch that companies are navigating the dual demands of innovation speed and safety. “You need to maintain momentum while ensuring security; these must coexist,” he stated.
According to Cisco’s 2024 AI Readiness Survey, a significant 71% of participants feel inadequately prepared to thwart unauthorized manipulations of AI within their organizations, while 67% emphasize their limited grasp of machine learning threats. Patel asserted that AI Defense aims to resolve these concerns effectively.
“Cisco AI Defense serves as a universal safety and security layer applicable across any AI model or application, regardless of the cloud environment,” Patel remarked.
Model Validation at Scale
The primary audience for Cisco AI Defense includes enterprise AppSecOps groups, providing them with mechanisms to validate AI models prior to their deployment in production settings.
Patel highlighted that the dynamic nature of AI models, with ongoing data adjustments, significantly alters the behavior of applications and agents. “As models evolve, the validation processes must also be continuously updated,” he explained.
In a strategic move to enhance its offerings, Cisco acquired Robust Intelligence in 2022, a startup founded by Harvard researchers Yaron Singer and Kojin Oshiba. This acquisition brings a core component to AI Defense: the Robust Intelligence Platform, which employs algorithmic red teaming techniques to detect vulnerabilities. Additionally, a novel method known as Tree of Attacks with Pruning allows for automation in testing large language models (LLMs).
Patel noted that Cisco AI Defense leverages detection models from Scale AI’s generative AI platform and integrates threat intelligence from Cisco’s Talos and the recently acquired Splunk to provide continuous model validation and automatic recommendation of necessary guardrails. These guardrails are designed to be distributed across the network fabric seamlessly.
“This provides a purpose-built model to assess whether an AI model will function as intended or could yield unexpected results,” Patel added, illustrating the efficiency of Cisco’s solution: while typical model validation can take weeks, the automated process can often complete in as little as 30 seconds.
An Industry-First?
Analysts speculate that Cisco might be among the first major companies to offer technology capable of such comprehensive automated model verification.
Frank Dickson, IDC’s group vice president for security and trust research, expressed skepticism about competitors offering similar capabilities, stating, “I haven’t encountered anyone else achieving something like this. The potential for automated pen testing in such a short timeframe is impressive.”
Scott Crawford, research director for 451 Research, concurs, noting that while various major vendors are tackling GenAI security from diverse perspectives, Cisco has made a distinctive move with its acquisition of Robust Intelligence, the centerpiece of this initiative. He mentioned that potential opportunities exist for other startups in this emerging market.
Given the rising incidence of attacks targeting vulnerable AI models, addressing AI security will be critical this year. Crawford pointed out early examples of LLM vulnerabilities being exploited, highlighting the ongoing discourse regarding the manipulation and security risks surrounding AI technologies.
Such attacks, often termed LLMjacking, often involve exploitation techniques like prompt injections, supply chain attacks, and data poisoning. A significant LLMjacking incident was reported last year by the Sysdig Threat Research Team, which tracked the theft of cloud credentials targeting ten cloud-hosted LLMs using a vulnerable Laravel version (CVE-2021-3129).
Source
www.darkreading.com