October 2025 Summaries

AI's integration into software development, particularly in code generation, has been noted for boosting productivity but also raises significant cybersecurity concerns, as a recent study reveals that iterative AI-generated code tends to become less secure. Despite using security-focused prompts, vulnerabilities increase with more iterations, highlighting issues like cryptographic library misuse and outdated security patterns. The study, which utilized OpenAI’s GPT for generating 400 code samples, found that vulnerabilities peaked in later iterations, even when secure outputs were requested. This suggests that while AI coding tools can enhance productivity, they necessitate robust security measures and guardrails to mitigate risks. Best practices such as spec-driven development, security-conscious prompts, and thorough code reviews are recommended to manage these risks effectively. The findings stress the importance of modernizing security approaches to align with AI advancements in software development without compromising on security.

The shift-left approach to application security, which aimed to catch vulnerabilities early in the development lifecycle to reduce costs and enhance security, has faced significant challenges due to outdated tools and methods. Traditional static application security testing (SAST) tools have been ineffective, creating alert fatigue with high false positive rates, poor developer experience due to security-first language, and misalignment between security and engineering teams. These issues have resulted in security theater rather than real risk reduction. A more effective strategy, termed "shift security down," suggests integrating security analysis directly into development platforms, emphasizing precision over volume, prioritizing exploitable risks through reachability and dataflow analysis, and providing developers with actionable evidence and guidance. This approach leverages modern application security platforms that utilize AI and policy as code to automatically triage findings, ensuring relevant issues are addressed efficiently and enhancing trust and effectiveness in security practices.

Static Application Security Testing (SAST) tools have faced criticism for their high false positive rates, with independent benchmarks indicating that traditional SAST tools produce a significant number of inaccuracies. Initially designed to make secure coding scalable, SAST tools have struggled to keep up with modern software development, often failing to recognize complex application structures and producing irrelevant results. The evolution of application security has been marked by a series of shifts that automate security measures, moving responsibility from humans to compilers, frameworks, and infrastructure. With the emergence of autonomous coding agents, the next progression aims to incorporate security best practices by default, allowing machines to generate and verify secure code, thereby further reducing human error. This ongoing evolution reflects a broader trend in application security towards automation and abstraction, optimizing systems to produce secure code with minimal human intervention, and suggests that SAST will need to adapt to maintain relevance in this new era.

Imperva recently revealed a high-severity vulnerability (CVE-2025-53967) in the popular Framelink Figma MCP server, which has since been patched. This flaw allowed for remote code execution (RCE) by exploiting a design oversight in the server's fallback mechanism, posing risks to developer machines and connected networks. The Framelink Figma MCP, a third-party plug-in enabling Figma tools in AI code agents, gained popularity due to its early release ahead of the official Figma MCP Server. The vulnerability affected versions before 0.6.3, allowing unauthorized users to execute system commands via crafted HTTP POST requests. Imperva's timeline detailed their disclosure process and the subsequent patch release. The incident highlights the importance of vetting MCP servers due to their potential security risks, including supply chain compromises and API abuses. While MCP servers improve code security by enriching model context, their development often lacks mature security assessments, emphasizing the need for thorough scanning and monitoring. Organizations must balance risk tolerance and innovation when considering third-party MCP server use, recognizing that independently maintained projects like Framelink Figma MCP may require additional diligence in areas such as dependency management and security updates.

In early 2023, Endor Labs introduced DroidGPT, marking its entry into integrating generative AI with application security, revealing a readiness among application security engineers to engage conversationally with their tools. This led to an expansion of AI capabilities in their platform, including AI Model Discovery and AI Security Code Review, prompting the company to question the traditional user experience of enterprise software. They explored transforming the experience from "click and configure" to "ask and collaborate" by conducting research with security engineers and AppSec leaders, who expressed a desire for proactive automation, actionable chat interactions, and personalized AI learning based on user patterns. This research informed the development of agentic chat, a next-generation conversational tool that integrates preset prompts, explanatory capabilities for package errors, and deeper insights from vulnerability searches to streamline workflows and enhance user trust. As Endor Labs continues to innovate, the focus remains on designing tools that enable AppSec engineers to secure software efficiently and intelligently.