Advanced AI models can increase cyber risk by making it faster and easier to identify security weaknesses and potential targets. This page brings together resources from Business Roundtable member companies – including technical tools, best practices, playbooks and training materials – to help businesses of all sizes strengthen cyber resilience.
Examines Anthropic's Mythos evaluation, highlighting how advanced AI models can accelerate cyberattack workflows and emphasizing the need for stronger defenses, governance, and preparedness as AI-enabled cyber capabilities continue to evolve.
Highlights survey findings that attackers are adopting AI faster than organizations are adapting, and warns that budgets, talent, technology maturity, and AI-based defensive capabilities lag behind the threat.
Suggests best practices for companies in response to the increased risk and speed of cyber threats, to include redesigned security operating models, clearer decision rights, defined workflows, accountability, and tighter coordination across business, IT, and security teams.
Introduces BCG’s Framework for Agentic AI Secure Transformation, focusing on the capabilities needed to make agents reliable, secure, and ready to move safely from proof of concept to broader enterprise use.
Describes Cisco’s Foundry Security Spec, highlighting how its open-source, model-agnostic specification provides roles, guardrails, validation workflows, and auditable evaluation structures to help defenders build AI-powered security assessment systems that produce bounded, verifiable findings. This specification is a model agnostic and tech stack neutral blueprint for building a model harness.
Describes Cisco’s Project CodeGuard framework, highlighting how its open-source, model-agnostic ruleset helps secure AI-generated code through secure-by-default guidance, agent-specific translators, automated validation, and defense-in-depth controls across the AI coding lifecycle.
Shares lessons from participation in Anthropic's Project Glasswing and OpenAI's Daybreak evaluations on how highly capable AI models could enhance cyberattack capabilities and what defenses organizations should implement in response. It provides practical guidance on strengthening cyber resilience through secure AI deployment, vulnerability management, governance, workforce readiness, and adoption of AI-assisted defensive capabilities.
Describes and categorizes the threat posed by “AI Agent Traps,” adversarial content designed to manipulate, deceive, or exploit visiting agents.
Describes the six core elements of Google’s Secure AI Framework: expand strong security foundations to the AI ecosystem; extend detection and response; automate defenses; harmonize platform-level controls to ensure consistent security across the organization; create faster feedback loops for AI deployment; and contextualize AI system risks.
Presents a framework for securing AI agents across three layers: individual-agent security, multi-agent ecosystem risks, and cyber-defender empowerment.
Introduces the Adversarial Robustness Toolbox (ART), an open-source Python library for machine learning security. ART enables red and blue teams to evaluate and defend AI models against the adversarial threats of evasion, poisoning, extraction and inference, and supports all popular machine learning frameworks, data types and tasks.
Introduces the Open Source AI Project Governance and Security Baseline, a framework that establishes security and governance standards for open source AI projects, building on Open Source Security Foundation principles and extending them with AI-specific requirements to help build trusted AI.
Outlines actions organizations can take to strengthen cyber resilience in preparation for increasingly capable AI systems, focusing on governance, identity security, software supply chain protection, data security, and AI-enabled defensive capabilities.
Introduces the Firewall Attack Detections and Extractions (FADE) dataset, an open-source tool that uses firewall rules to generate large-scale network attack mutations to help expose obfuscation techniques, improve detection testing, and support stronger AI-driven cybersecurity defenses.
Introduces MCPSafetyScanner, an open-source multi-agent tool for auditing Model Context Protocol (MCP) servers, identifying exploitable weaknesses from MCP tools and resources, researching related remediations, and generating safety reports to help developers harden agentic AI workflows before deployment.
Explains how machine learning operations, or MLOps, enables secure and reliable government AI by using repeatable deployment pipelines, continuous monitoring, governance controls, model retraining, and operational oversight to maintain trusted AI systems at scale.
Outlines a three-phase framework of assessment, protection and platformization to help organizations defend against AI-driven threats. Draws on early testing of frontier AI models to explain how attackers will accelerate vulnerability discovery, exploit generation and attack cycles and why defenses must operate at machine speed.
Shares how frontier AI models act as full-spectrum security researchers, enabling autonomous zero-day discovery, faster exploitation of known vulnerabilities and complex exploit chaining. Recommended actions and guidance for security teams include assuming breach conditions, implementing strong software supply chain governance, accelerating patching and automating incident response.
CEO Nikesh Arora argues that frontier AI models represent a structural shift in the threat landscape by making sophisticated attack capabilities broadly accessible, and outlines what effective defense requires, including comprehensive sensor coverage, context-rich security data, consolidation of fragmented security tools and responsible release practices by AI labs.
Describes how advances in frontier AI are transforming cybersecurity—enabling rapid, automated discovery and exploitation of vulnerabilities—and how Visa is using AI-driven tools and architectures to strengthen cyber resilience and respond at machine speed.
Describes the Visa Vulnerability Agentic Harness (VVAH), an open-source enterprise orchestration framework built for autonomous cybersecurity vulnerability discovery and remediation. Stemming from Anthropic's Project Glasswing, VVAH automates the pipeline from finding security flaws to deploying validated fixes.
