Photo credit: www.csoonline.com
Technical, organizational, and cultural challenges have hindered enterprises from adequately addressing vulnerabilities identified during penetration testing, a situation that generative AI seems to worsen rather than alleviate.
A report by Cobalt, a provider of penetration testing as a service, reveals that organizations remediate less than half of discoverable vulnerabilities (48%), and this figure drops to a mere 21% for flaws associated with generative AI applications.
Vulnerabilities classified as either high or critical in severity tend to have a higher resolution rate, standing at 69%. This increase highlights the importance of addressing severe vulnerabilities as a priority.
Notably, the median duration to resolve significant vulnerabilities has significantly decreased from 112 days in 2017 to just 37 days last year, demonstrating the positive effects of “shift left” security initiatives as noted by Cobalt.
Patching Headaches
Organizations sometimes opt to accept certain risks instead of incurring the operational disruptions or costs associated with patching vulnerabilities.
Limited resources and inadequate remediation strategies contribute to the slow pace of patching efforts. Additionally, challenges arise when vulnerabilities are located within legacy systems that prove difficult to update or replace.
Cobalt’s researchers pointed out that “Some organizations do only what they’re legally obligated to do for compliance or third-party approval — merely conducting a pentest. Addressing risk through remediation takes a back seat.” They emphasize that the issue largely stems from a complex blend of organizational dynamics involving people, processes, and technology.
Next Gen-AI-eration
The latest iteration of Cobalt’s State of Pentesting Report indicates that a significant number of firms have engaged in penetration testing of large language model (LLM) applications, with around 32% of these tests uncovering serious vulnerabilities.
Common issues such as prompt injection, model manipulation, and data leakage were identified, yet only 21% of these weaknesses were addressed. Cobalt warns that AI development is advancing rapidly without adequate safeguards.
The study is based on data from over 5,000 penetration tests conducted by Cobalt. A concurrent survey among customers revealed that more than half (52%) of security leaders feel pressured to prioritize speed over security measures.
Vulnerabilities ‘Flagged but Not Fixed’
Independent security experts corroborate Cobalt’s findings, noting a general trend where organizations lag in addressing known vulnerabilities.
James Lei, a seasoned engineering executive now serving as COO at legal services company Sparrow, remarked, “Organizations are typically slow to address known vulnerabilities, and this is seldom due to unawareness. These issues are identified but left unresolved.”
The delay in addressing vulnerabilities is attributed to competing business priorities.
“Security teams are often overburdened, and engineering teams prioritize feature deployment. Unless there’s regulatory pressure or an experienced breach, a ‘known issue’ generally doesn’t receive the urgency it requires,” Lei added.
Bug Remediation in the Age of AI
Generative AI applications introduce unique challenges that complicate the process of fixing vulnerabilities.
“Many of these applications are developed quickly, using emerging frameworks and third-party tools that lack extensive testing in real-world environments,” said Lei. “They create unfamiliar attack vectors, unpredictable model behavior, and dependencies that teams may not fully understand.”
Resolving these vulnerabilities can be intricate and time-consuming, requiring in-house expertise that may not always be available.
A generative AI application typically consists of two main components: the application and the generative AI, often represented by an LLM like ChatGPT.
Inti De Ceukelaire, chief hacker officer at bug bounty platform Intigriti, explained, “Fixing traditional application vulnerabilities is no different from addressing conventional bugs, but LLM vulnerabilities present unique challenges.”
For instance, if a generative AI app’s functionality for retrieving documents contains a flaw, developers can amend the code directly. Conversely, a flaw within the LLM itself is inherently more difficult to resolve due to the opacity of the model’s operational behavior.
“Assumptions can be made to adjust the model, but one cannot guarantee that adjustments effectively resolve the issue,” De Ceukelaire cautioned. “Drawing parallels with standard patching practices may not be entirely appropriate.”
In response to Intigriti’s observations, Cobalt clarified that their explorations into generative AI primarily focus on “validating the integrity of LLM-supported systems rather than evaluating the comprehensive behavior or outputs of the LLM itself.”
Bug Triage
For Chief Information Security Officers (CISOs) looking to enhance remediation rates, the key lies in simplifying prioritization of security fixes. This approach may involve integrating security tools earlier in the development lifecycle and establishing performance indicators focused on the timely resolution of serious threats.
“It also requires clear ownership — appointing individuals responsible for ensuring that vulnerabilities are rectified rather than merely documented,” emphasized Lei from Sparrow.
Experts recommend that security practitioners concentrate their limited resources on the most critical vulnerabilities, particularly those exposed directly to the internet.
Moreover, addressing accidental exposures and minimizing technical debt should also take precedence, according to Tod Beardsley, VP of security research at runZero.
“An effective penetration test will guide CISOs in pinpointing areas where threats may flourish instead of simply detailing a list of critical vulnerabilities devoid of context,” Beardsley explained.
Security teams often find themselves overwhelmed by the volume of vulnerabilities reported by both pentesting and scanning tools.
“The amount of information can be daunting, making it challenging for teams to prioritize remediation efforts based on risk severity,” pointed out Thomas Richards, infrastructure security practice director at Black Duck.
Richards echoed Beardsley’s sentiments, asserting that it is essential to interpret the outcomes of penetration tests in a proper context.
“After receiving a penetration test report, internal security teams take time to assess its accuracy and determine subsequent actions,” he stated. “This step, although time-consuming, allows organizations to focus on mitigating the highest risks first.”
Vulnerability scanning results require even more scrutiny.
“Through our automated tools, we often discover the default severity ratings are misleading when considering variables like exploit availability, network accessibility, and other remediation efforts that mitigate risk,” Richards elaborated. “Frequently, critical issues may have already been addressed on high-risk systems.”
Source
www.csoonline.com