{
  "id": 8948414,
  "title": "AI Security Is an Engineering Problem — How to Solve It at Every Layer of the Agent Stack",
  "url": "https://urgent.news/2026/09/21/ai-security-is-an-engineering-problem-how-to-solve-it-at-every-layer",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-21T14:51:34.000Z",
  "source": {
    "name": "NVIDIA Blog",
    "slug": "nvidia-blog",
    "url": "https://blogs.nvidia.com/blog/ai-security-agent-stack/"
  },
  "original_language": "en",
  "account": "AI security is fundamentally an engineering challenge. To address this, companies must establish clear security requirements, implement enforceable controls, assign ownership and demonstrate that protective measures function effectively. As AI technology evolves, the industry must accelerate security engineering, make defensive tools more accessible and disseminate successful practices more swiftly.\n\nThe principles that governed security in the early days of the internet and cloud computing remain relevant today. The introduction of AI agents, however, brings new capabilities that necessitate the adaptation of these principles to new operating conditions. These capabilities include reasoning, utilizing tools, and adjusting actions based on encountered data. Consequently, organizations face pressure to balance the benefits of AI, such as productivity gains, while simultaneously developing and refining practices to govern and secure AI systems.\n\nApplications depend on various components, including code, data, identities, services, and infrastructure. Security in this context is contingent upon the interaction among these components. AI agents extend the system by adding new functionalities. Each component of the system requires security measures, and comprehensive protection necessitates implementing controls at each layer as data, instructions, and actions traverse the system.\n\nConsider an AI agent updating a customer record. If the agent encounters malicious instructions within an attached document, it may attempt to export customer data to an unauthorized destination. A network policy should prevent such data transfer. Additionally, protected logs should document the attempted tool call, authorization decision, and the outcome, enabling the security team to identify the tool involved and the destination it intended to access. The permission to modify a customer record should not inherently grant the right to export associated data.\n\nWhile instructions and safeguards can guide agent behavior, enforceable boundaries are equally essential. Each agent must possess a verifiable identity and credentials restricted to its designated task. Organizations must formulate clear policies outlining the information agents can access, the systems they can alter, and the actions that require approval. Within these defined boundaries, consequential actions and permission changes must still undergo human approval. Additionally, teams must verify the source and integrity of the tools, skills, and dependencies utilized by agents. Upon encountering issues, protected records of tool calls, authorization decisions, and outcomes facilitate investigators in reconstructing the incident. Clear procedures for revoking access and containing incidents ensure that this evidence can be effectively utilized.\n\nNVIDIA's OpenShell is an open-source, secure runtime that enforces policies beyond the agent's reach and governs how agents interact with data, network, and system resources. This runtime facilitates sandboxed execution. Open Secure AI Alliance partners, including Cisco and JFrog, are building upon OpenShell. Cisco's DefenseClaw introduces a governance layer, while JFrog integrates with OpenShell to scan and validate agent skills, enforcing policies on which skills agents can access.\n\nEngineering teams require evidence that security measures effectively block unauthorized access attempts or the transmission of sensitive data to unauthorized destinations. Testing should encompass scenarios such as unauthorized credential acquisition, attempts to alter permissions, interference with monitoring, and should be repeated following significant changes to models, tools, or workflows. Assigning named owners to these tests ensures that failures are investigated, addressed, and incorporated into repeatable tests, enabling teams to verify that the implemented fixes maintain their effectiveness in subsequent releases.\n\nExamples of such testing tools include CrowdStrike's SafeMind, designed for testing and enhancing defenses through repeated attack simulations, and Palo Alto Networks Prisma AIRS, which facilitates continuous red-teaming as models and applications undergo modifications. Defenders require appropriate tools suited to their specific tasks, data, and environment. Closed models provide managed capabilities and services, while open models offer defenders the opportunity to inspect critical components, adapt strategies, and operate within infrastructure they control. During incident response, these capabilities enable teams to reproduce failures, test fixes against their own systems, and maintain sensitive evidence within their environment.\n\nAI-powered tools, such as Capital One's VulnHunter for AI-powered code security and ReversingLabs' Spectra Assure for AI-powered analysis of software packages to detect malware and tampering, can bolster these efforts. Their effectiveness should be evaluated through reproducible findings, verifiable fixes, and accelerated response times. Open research and the sharing of tools and expertise can significantly enhance the defensive posture of organizations. NVIDIA's security research and the Open Secure AI Alliance contribute to this knowledge exchange by providing research, practical tools, and industry expertise to the broader security community. Ultimately, AI security is an engineering problem that requires defined security requirements, enforceable controls, accountable ownership, and verifiable evidence that protective measures operate effectively.",
  "summary": "AI security is an engineering problem. That means defined security requirements, enforceable controls, named owners and evidence that protections work. As AI becomes more capable, the industry must accelerate security engineering, broaden access to defensive tools and share what works faster. Technology Changes, Security Fundamentals Endure The internet and cloud computing changed how software…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}