August 2025 Summaries

In August 2025, several npm packages associated with the Nx build system were compromised by a malware that exploited AI code assistants to collect sensitive information, such as SSH keys and API tokens, and upload them to GitHub repositories. This incident highlights a new trend where attackers leverage the capabilities of AI assistants like Claude, Gemini, and Amazon Q to autonomously search for sensitive data without implementing their own search logic. The malicious packages, which included a script called "telemetry.js" executed post-installation, were swiftly removed from npm following their discovery by vigilant GitHub users and project maintainers. The attackers took advantage of a compromised npm token with publish rights, a known vector in software supply chain attacks, and utilized AI assistants' "hands-free" mode to bypass permission requirements. Although some AI models, like Claude Code, refused to execute the malicious requests, the incident underscores the potential for abuse when these models are misaligned, as well as the importance of prompt detection and response in mitigating such threats.

Software Composition Analysis (SCA) faces challenges in identifying dependencies for older languages like C and C++ due to the prevalent practice of developers copying third-party code into repositories without maintaining precise logs. This makes creating an accurate Software Bill of Materials (SBOM) difficult, increasing security and licensing risks. Endor Labs addresses these challenges by developing an innovative approach to SCA, leveraging techniques from code clone detection and semantic search to build a comprehensive index of C/C++ libraries. They utilize a combination of hash signatures and text embeddings to identify the origins of cloned source code, even creating an ingestion infrastructure to index various archives beyond GitHub. Through a combination of semi-automated analysis and manual annotation, Endor Labs' approach enables precise identification of library versions, surprising users with their depth of detection and potentially revealing unknown library usages. Their method has been benchmarked against high-profile OSS projects, demonstrating effective performance in library identification compared to competitors in the C/C++ SCA space.

CodeRabbit, a startup offering an AI-powered code review tool, faced a security breach when researchers discovered a vulnerability in the Rubocop tool configuration that allowed the execution of malicious code, exposing sensitive environment variables including the CodeRabbit GitHub App private key. This incident highlighted a significant software supply chain risk as the compromised credentials could allow attackers to perform unauthorized actions on any installation. Despite the potential for widespread impact, CodeRabbit quickly responded by disabling the vulnerable tool, rotating secrets, and implementing sandboxing measures. This situation underscores the importance of robust security practices, such as zero trust architecture, least privilege principles, and defense in depth, to mitigate risks associated with executing user-controlled code in privileged environments. The incident serves as a reminder for both vendors and users of GitHub Apps to ensure architectural safeguards are in place to limit potential damage from security breaches.

Greg Pettengill, a Principal Product Security Engineer at Five9 with over 30 years in IT and 20 years in security, shares his insights on selecting security vendors, particularly favoring early-stage startups for their responsiveness and innovation. He emphasizes the importance of thoroughly understanding the problem and technology needs, akin to creating a job description, before selecting a vendor. Pettengill's strategy involves a detailed vetting process similar to a job interview, focusing on both technical capabilities and cultural fit. He values transparency and verifiability in technical discussions and seeks vendors that demonstrate strong communication and partnership potential. His approach has led to successful vendor relationships and mitigated the risks associated with being an early adopter of new technologies.

Five9, a leading contact center software provider, faced challenges in managing third-party library vulnerabilities within its applications, as their existing Software Composition Analysis (SCA) tool failed to reduce risks effectively, leading to a backlog of issues and potential exposure affecting their uptime standards. To address this, Five9 chose Endor Labs, an AppSec platform, for its ability to accurately prioritize vulnerabilities based on reachability, reducing false positives and enabling efficient vulnerability management with minimal resources. Endor Labs' willingness to partner and adapt to Five9's needs, along with its robust CI/CD integration and automation capabilities, allowed the company to transform its software supply chain security program, achieve a 96.5% noise reduction, and respond rapidly to zero-day vulnerabilities. The platform's approach to security, including proactive measures against AI risks, has empowered Five9 to systematically improve code quality and maintain high security standards, with a single security engineer effectively supporting over 250 developers.

Cursor, a rapidly growing product owned by Anysphere, utilized Endor Labs for software composition analysis (SCA) to enhance security while maintaining developer productivity. The existing SCA tool was ineffective due to excessive noise and risky upgrade advice, leading Cursor to seek a more precise solution. Endor Labs' function-level reachability analysis significantly reduced irrelevant findings by 97.5%, allowing the security team to focus on impactful upgrades and avoid system instability. The integration with Endor Labs streamlined vulnerability management, enabling the security team to manage dependency upgrades and maintain stable code while allowing engineers to focus on product development.

Machine learning (ML) and large language models (LLMs) are increasingly integral to modern application development, but their integration poses significant security risks, especially when adopted without oversight, a phenomenon known as "shadow AI." This occurs when developers integrate AI models and services into projects without formal review, leading to potential vulnerabilities such as prompt injection, which can be exploited to manipulate outputs or gain control over environments. Organizations often lack clear guidelines for AI use, resulting in challenges in auditing and ensuring compliance with regulations like GDPR. To mitigate these risks, firms need robust governance frameworks to manage AI usage, including policy frameworks, visibility into AI components, and automated discovery tools. Endor Labs offers solutions to help organizations build inventories of AI components, assess supply chain risks, and establish policies for safe AI integration, thereby enabling effective governance and compliance.

Shipping containers with end-of-life (EOL) operating systems or packages pose a significant security risk as they no longer receive security patches and often lack upstream visibility into new vulnerabilities. Endor Labs addresses this issue by supporting automated detection of EOL components in container images, providing visibility that allows teams to identify unsupported software early in development or CI pipelines. This detection is powered by lifecycle metadata from public sources, covering major distributions like Ubuntu, Debian, Alpine Linux, and Red Hat Enterprise Linux, with policies that can enforce actions such as failing builds if unsupported components are detected. Endor Labs' approach integrates EOL detection as an essential part of container scanning, enabling application security teams to secure the software supply chain by eliminating EOL software before deployment.

Recent academic research highlights that over 40% of AI-generated code solutions, created using large language models (LLMs), contain security vulnerabilities. These issues, although not new, exhibit novel patterns, occurring in unexpected ways and often bypassing existing security measures. LLMs are trained on diverse open-source code, which includes both high-quality examples and flawed or insecure code snippets. This mixed training data leads to the models replicating prevalent insecure coding patterns, such as missing input validation and injection flaws, which align with the CWE Top 25 security weaknesses. Prompts lacking explicit security guidance often result in code with authentication and authorization failures, such as broken access controls and hard-coded credentials. These vulnerabilities are concerning, especially as AI is increasingly used to develop full-stack applications, where human oversight may be limited.

Engineering organizations are increasingly adopting AI coding assistants like GitHub Copilot to enhance productivity, with claims of significantly faster coding times, yet the anticipated increase in engineering velocity has not matched the volume of code generated. Sundar Pichai, CEO of Alphabet, highlighted that while 30% of new code at Google is AI-generated, the actual increase in engineering velocity is around 10%. This discrepancy is attributed to the unchanged pace of code review processes, which remain a bottleneck despite the accelerated code generation. Studies show that while AI tools boost productivity, the manual code review process, essential for addressing security and correctness issues, has not scaled accordingly, often requiring significant time and attention from senior engineers. Automated secure code review platforms like Endor Labs have emerged to address this challenge, using advanced systems to perform real-time analysis and streamline the review process, dramatically reducing the need for manual reviews and enabling faster, more secure code delivery. By efficiently prioritizing and addressing potential risks, such platforms help engineering teams balance speed with security, thereby unlocking substantial engineering hours and improving delivery timelines.

In a discussion hosted by Jenn Gile of Endor Labs, Jeremiah Kung from AppLovin and Aman Sirohi from People.ai explore the transformative role of AI in application security and developer velocity, emphasizing its potential to make the "shift-left dream" a reality by integrating AI code assistants. The panelists highlight the importance of moving away from noisy security tools to high-signal solutions, fostering a partnership between developers and security teams for enhanced trust and efficiency. They share insights into the adoption of AI tools like Endor Labs, which streamline security processes and improve application release speed, while acknowledging challenges in rebuilding developer trust after previous tool failures. The conversation also touches on the excitement and innovation in the security industry, as evidenced by the dynamic environment at the Black Hat conference, where collaboration and influence over emerging technologies are key themes.

AI-native applications, designed to utilize the capabilities of large language models and AI technologies, present unique security challenges, notably prompt injection attacks. These attacks can manipulate AI inputs to produce harmful outputs, posing risks to systems relying on language models and ranking high in OWASP's 2025 security concerns. An example is a Python application where despite efforts to redact sensitive data, the LLM could still be manipulated, showing limitations of traditional security tools. Endor Labs addresses these vulnerabilities with its AI Security Code Review that analyzes code changes, categorizes security risks, and provides detailed feedback to developers and security teams. This system helps identify complex vulnerabilities and ensure security measures like prompt sanitization and output filtering are in place, which is crucial as 62% of AI-generated solutions may contain design or security flaws.

During a session at Black Hat 2025, Dr. Jasyn Voshell, who leads product security at Zebra Technologies, discussed the company's approach to product security and the integration of various tools to manage vulnerabilities, highlighting the importance of building trust and efficiency in development teams. Zebra Technologies, initially known for its barcode innovations, now plays a significant role in various industries through its RFID technology, which tracks everything from sports players' movements to the supply chain of grocery items. Voshell emphasized the importance of "shift left" practices in software development to address security vulnerabilities early in the process. The adoption of Endor Labs and other security tools like Snyk and Nucleus has helped Zebra reduce noise in vulnerability detection and improve developer confidence by focusing on reachability—determining whether a vulnerability is actively exploitable. The talk also touched on the challenges of integrating new security solutions within existing systems and the necessity of collaborating with sales teams to communicate security measures effectively to clients, particularly when addressing concerns about potential vulnerabilities. Voshell underscored the ongoing need for innovation and adaptability in security strategies, particularly in the face of evolving technologies such as AI, to maintain product integrity and customer trust.

Experimentation with large language models (LLMs) like GPT-4 and Claude 4 Sonnet for personal projects has revealed that these models often import outdated open-source dependencies, some of which contain known vulnerabilities. This issue arises because LLMs rely on past training data and are not constantly updated with real-time information unless specifically connected to the internet and equipped with tools to fetch the latest data. A notable example includes the Axios library, where a critical vulnerability was not recognized by the model due to its training cutoff date. To address these challenges, developers and security teams are encouraged to enhance their security awareness by prompting models to use the latest versions, implementing rules for dependency checks in their IDEs, and utilizing security tools like the Endor Labs MCP Server to provide up-to-date security context. These measures ensure that code generated by LLMs is secure, up-to-date, and does not inadvertently introduce vulnerabilities into software projects.

AI coding agents are becoming integral to modern software development, capable of quickly generating boilerplate code and complex integrations, but they often produce insecure code due to training on publicly available data containing both secure and insecure examples. The anti-pattern avoidance prompt pattern is a zero-shot technique that addresses this issue by instructing AI models to generate code while avoiding specific security weaknesses identified by Common Weakness Enumeration (CWE), such as CWE-20 for improper input validation. This approach, which requires more security knowledge than other methods like the secure vs insecure diff pattern, involves identifying relevant CWEs for the task at hand and instructing the language model to avoid these weaknesses, thereby enhancing the security of AI-generated code by focusing on secure implementation alongside functionality.