What is Frontier AI Security? Definition, Threats, and Defenses for 2026

What is Frontier AI Security?

Frontier AI security is the practice of defending organizations against attackers who are using the most capable and advanced AI models to accelerate, automate, and scale their offensive operations. It is distinct from securing your own AI systems: frontier AI security focuses on the external threat posed by adversaries who wield cutting-edge models to discover vulnerabilities, execute attacks, and evade detection faster than traditional defenses can respond.

This guide covers what frontier AI is, how it differs from conventional AI security, what Mythos-class attacks look like in practice, how security platforms defend against these threats, and the five-step framework security teams and chief information security officers can adopt today. Bitsight's research and intelligence capabilities are woven throughout to show how defenders can operationalize each step. For a deeper CISO readiness roadmap, see our companion guide and a detailed platform comparison.

What is frontier AI?

Frontier AI refers to the most advanced and capable AI systems available at any given time, models that sit at the outer boundary of what technology can currently do. As defined by the UK National Cyber Security Centre (NCSC), these systems can perform complex tasks that generate both significant benefits and serious risks to cybersecurity. Frontier models are not incremental improvements over prior generations; they represent step-change gains in reasoning, code generation, autonomous planning, and multi-step task execution.

The Canadian Centre for Cyber Security (CCCS) describes frontier models as highly capable AI systems that can read and generate code, automate vulnerability discovery, craft sophisticated phishing campaigns, and orchestrate multi-stage attacks faster than traditional offensive methods. For defenders, this matters because it means capabilities once exclusive to nation-state threat actors are becoming accessible to a much wider range of adversaries, including those with limited prior technical expertise.

Is frontier AI security the same as AI security?

No, and the distinction is important. The term "AI security" is often used to describe the practice of securing your own AI systems: protecting large language model deployments, preventing data poisoning, hardening AI pipelines, and governing model access. Frontier AI security, as covered on this page, addresses the opposite direction of risk: defending your organization against attackers who are actively using frontier AI models as offensive weapons.

These are related but separate disciplines. Exposed AI infrastructure, such as publicly accessible AI orchestration platforms, model serving endpoints, and AI coding tools, may appear in this guide as attack surface that adversaries exploit. But the core subject here is not AI governance or AI compliance; it is how to detect, prioritize, and respond to threats that are being materially amplified by the most capable models on the market.

What is a Mythos-class attack?

A Mythos-class attack refers to a cyber operation that is enabled or substantially accelerated by a frontier AI model capable of autonomous, multi-step offensive behavior. The term takes its name from Anthropic's Claude Mythos Preview, which the UK AI Security Institute (AISI) evaluated in April 2026. In controlled evaluations where Mythos Preview was explicitly directed and given network access, the AISI found that it could execute multi-stage attacks on vulnerable networks and discover and exploit vulnerabilities autonomously, tasks that would take human professionals days of work.

The AISI's evaluation found that frontier AI capabilities in cyber offense have been advancing rapidly, with the length of tasks that AI models can autonomously complete in cyber challenge environments doubling every few months. Anthropic did not release Mythos commercially, citing that frontier models had reached a point where they could surpass all but the most skilled humans at finding and exploiting software vulnerabilities.

For practical purposes, a Mythos-class attack shares several defining characteristics:

• Machine-speed exploitation: Frontier models identify and adapt to vulnerable conditions in seconds or minutes, collapsing the window between discovery and active exploitation. Bitsight Threat Intelligence (TI) observed that frontier AI models are compressing the time between exposure and exploitation and helping attackers identify which vulnerabilities are relevant faster than ever.
• Autonomous reconnaissance: Rather than requiring a human operator to enumerate targets, frontier AI agents can autonomously scan, fingerprint, and prioritize attack paths across large external attack surfaces, removing the defender's advantage of private discovery.
• Exploit chaining: Frontier models can combine multiple low- and medium-severity weaknesses into a high-impact compromise path. The CCCS has noted that AI enables threat actors to find and combine weaknesses to carry out attacks through what is known as vulnerability chaining.
• Collapsed patch windows: Public proof-of-concept (PoC) code and AI-assisted exploit adaptation are compressing the disclosure-to-exploitation window. According to the Bitsight's State of the Underground 2026, critical vulnerabilities can no longer be patched on a schedule. The European Union Agency for Cybersecurity (ENISA) found that the gap between a vulnerability being discovered and being exploited has shrunk from years to months and, in some cases, to minutes.

What are the types of frontier AI attacks?

Bitsight Threat Intelligence tracks the evolution of AI abuse across open, deep, and dark web sources. The following attack types represent the clearest patterns observed during the July 2025 to July 2026 period, as documented in Bitsight's report From Jailbreaks to Agentic Attacks: The Evolution of AI Abuse.

• Operationalized jailbreak workflows: Jailbreaking is when actors exploit flaws in AI systems to remove or limit guardrails, usually to leverage the system for harmful actions like ransomware and malware development. Bitsight TI observed that jailbreaks have evolved far beyond simple copy-and-paste prompts: they now exist as packaged workflows with mode labels such as GODMODE (also stylized as G0DM0D3), PARSELTONGUE, and ULTRAPLINIAN, using obfuscation, multi-model routing, and retry logic across Claude, GPT, Gemini, Llama, and Grok. In 2025, Bitsight TI observed approximately 5.1 million underground mentions of Gemini, 1.4 million of ChatGPT, 697,000 of Grok, and 656,000 of Claude, alongside extensive discussion of how to circumvent each platform's guardrails.
• AI-run attack infrastructure: Frontier models are being integrated directly into attacker workflows to handle tasks that previously required significant human labor. Bitsight TI observed a Russian-speaking actor who used a jailbroken AI coding command-line interface (CLI) to complete a full command-and-control (C2) migration in approximately six minutes while performing roughly 11% of the work manually. The model handled the remaining technical execution autonomously.
• Agentic execution via indirect prompt injection: AI coding agents can be weaponized through repositories that appear benign. Bitsight TI documented a proof of concept in which a normal-looking GitHub repository led an AI coding agent to open a reverse shell through indirect prompt injection, creating a compromise path that required no conventional malware. The JADEPUFFER campaign demonstrated this risk at production scale: an agentic ransomware operation that entered through an internet-facing Langflow instance via CVE-2025-3248 and pivoted directly to a production database.
• Model Context Protocol tool poisoning: MCP is a protocol that allows AI agents to connect with external tools, files, APIs, and system resources. Bitsight TI found that MCP servers can expose files, environment variables, internal APIs, and shell-backed tools to AI agents. Tool poisoning and prompt injection can turn that access into a full compromise path, making MCP-connected environments a meaningful new attack surface category.
• AI assets as extortion targets: Proprietary AI models, training datasets, and source code have emerged as high-value theft targets. Bitsight TI observed a claimed leak that included proprietary AI models, datasets, and source code, an indication that AI intellectual property is increasingly treated as a leverage point in extortion operations.
• Exposed AI infrastructure exploited at scale: Bitsight TI observed a 360% increase in exposed AI-related tooling in 2025, reaching more than one million exposed services. This surface area gives attackers an expanding pool of entry points through unsecured AI deployment endpoints, orchestration platforms, and model APIs.

Why are frontier AI attacks harder to stop?

Traditional defenses were designed around a threat model in which human attackers operated at human speed. Frontier AI breaks that assumption in several ways that compound each other.

First, the velocity of exploitation eliminates the gap that defenders historically relied on for patching and containment. When a critical vulnerability is disclosed today, AI-assisted exploit adaptation can compress the window from weeks to hours or even minutes. The Bitsight State of the Underground 2026 found that all 10 highest-impact Known Exploited Vulnerabilities (KEVs) of 2025 carried Bitsight DVE (Distributed Vulnerability Exploitation) and CVSS (Common Vulnerability Scoring System) scores above 9, meaning the most dangerous vulnerabilities were the ones moving fastest toward active exploitation.

Second, autonomous reconnaissance removes the signal defenders once relied on to detect intruders early. When an AI agent can enumerate an attack surface, identify the most exploitable path, and begin execution without human pauses between steps, the behavioral signatures that trigger conventional detection tools become harder to observe in time.

Third, the operator threshold has dropped. The NCSC has noted that frontier AI makes it easier, faster, and cheaper for even low-skilled actors to carry out sophisticated attacks. Capabilities previously reserved for nation-state-level operators are now accessible to a broader range of threat actors. The Hexagonal Rodent campaign, a 2026 operation closely linked to Famous Chollima that targeted Web3 developers with malware including BeaverTail, OtterCookie, and InvisibleFerret, illustrated how frontier AI tools such as Cursor and ChatGPT are being integrated alongside conventional malware in active campaigns.

Fourth, the attack surface itself is expanding. Exposed AI-related tooling grew 360% in 2025, and many organizations have deployed AI services, orchestration layers, and MCP-connected tools without fully inventorying them as part of their external attack surface. Attackers are mapping this surface systematically.

Enisa's July 2026 position paper put it plainly: security fundamentals have not changed, they are simply being stress-tested at a new speed.

Real-world examples of frontier AI threats

The following examples draw exclusively on findings from official government and standards bodies and Bitsight research.

• UK AISI evaluation of Claude Mythos Preview (April 2026): The AI Security Institute (AISI) evaluated Anthropic's Claude Mythos Preview and found it to be the first model to complete an AISI cyber range end-to-end. In controlled conditions, it executed multi-stage attacks and discovered and exploited vulnerabilities autonomously. The AISI noted that these evaluations lacked active defenders and defensive tooling, which does not reflect well-hardened enterprise environments, but the capability baseline was nonetheless unprecedented.
• NYDFS frontier AI advisory (May 2026): On May 21, 2026, the New York Department of Financial Services (NYDFS) issued an industry letter to CISOs of regulated entities warning of heightened cybersecurity risks associated with frontier AI models that amplify the potency, scale, and speed of identifying vulnerabilities and exploits in information systems. The NYDFS directed regulated entities to reassess vulnerability management timelines, map third-party dependencies, apply additional testing to AI-generated code, and evaluate whether existing logging and alerting capabilities can keep pace with AI-enabled attack cadences.
• ENISA frontier AI position paper (July 2026): ENISA published its view on cybersecurity in the frontier AI era, finding that frontier AI models are challenging traditional security paradigms by compressing the vulnerability management lifecycle and attack chain from discovery to exploitation. ENISA described an "authority gap" in which human change advisory boards cannot approve a fix in the few minutes now available to counter an autonomous exploitation attempt.
• Canadian Centre for Cyber Security statement (June 2026): The CCCS warned that frontier AI models are rapidly transforming the cyber threat landscape, reducing the time organizations have to detect, contain, and respond to attacks. The CCCS noted that organizations should assume AI-driven exploitation may bypass preventative controls and significantly outpace vendors' capacity to publish and deploy corrective measures.
• JADEPUFFER agentic ransomware: Bitsight Threat Intelligence observed the JADEPUFFER campaign, in which an agentic ransomware operation entered through an internet-facing Langflow instance via CVE-2025-3248 and pivoted to a production database. The attack required minimal human direction at execution time.
• C2 migration in six minutes: Bitsight TI observed a Russian-speaking actor use a jailbroken AI coding CLI to complete a full C2 migration in approximately six minutes, performing roughly 11% of the work manually. The remaining execution was handled autonomously by the AI model.

How do security platforms defend against frontier AI?

The challenge frontier AI poses to security teams is not simply one of adding new tools. It requires a structural shift in how organizations collect intelligence, prioritize risk, monitor exposure, and coordinate across their vendor ecosystems. The following five-step defense framework, drawn from regulatory guidance and Bitsight research, provides a practical structure for that shift.

A five-step defense framework against frontier AI threats

Step 1: Inventory and continuously monitor your entire external attack surface.

You can't defend what you can't see. Frontier AI attackers begin by mapping your exposure systematically, and defenders must do the same. This means maintaining a continuous, accurate inventory of all internet-facing assets, including AI services, orchestration endpoints, MCP-connected tools, and third-party integrations, not just traditional infrastructure. Static point-in-time assessments are insufficient when the exposure landscape changes daily.

• Bitsight capability: External attack surface management (EASM) provides continuous discovery and monitoring of your full digital footprint, including cloud, shadow IT, and AI-related services, so security teams always have an accurate picture of what is exposed to the internet.

Step 2: Prioritize vulnerabilities by exploitation likelihood, not just severity score.

Frontier AI models help attackers rapidly identify which vulnerabilities are most likely to be successfully exploited. Defenders must move beyond static CVSS scores and prioritize remediation based on real-world exploitation intelligence. The Bitsight State of the Underground 2026 found that all 10 highest-impact 2025 KEVs carried Bitsight DVE and CVSS scores above 9, demonstrating that the most dangerous vulnerabilities are identifiable in advance when the right intelligence is applied.

• Bitsight capability: DVE (Distributed Vulnerability Exploitation) Intelligence combines real-world exploitation data and threat actor behavior to help security teams focus remediation on the vulnerabilities most likely to be targeted, rather than those with the highest paper score.

Step 3: Monitor underground and adversarial AI activity for early warning.

Frontier AI threats do not emerge without signals. Jailbreak toolkits, AI-assisted exploit frameworks, and operational tradecraft are discussed and distributed across underground forums, Telegram channels, GitHub repositories, and invite-only communities before they are deployed at scale. Intelligence collected from these sources gives defenders advance warning of the techniques, targets, and tools that attackers are preparing to use.

• Bitsight capability: Cyber threat intelligence processes more than 540 billion cyber events, draws on over 1,000 underground sources, and delivers 7 million intelligence items daily, giving security teams the adversarial context they need to anticipate frontier AI-enabled attacks before they land.

Step 4: Continuously measure and improve your own security posture.

Regulatory bodies including NYDFS and the CCCS are explicit that organizations must raise their security baselines in response to frontier AI. This means continuously measuring controls effectiveness, ensuring patching cadences have accelerated to match compressed exploitation windows, and validating that logging and alerting capabilities can keep pace with machine-speed attack patterns. ENISA's guidance frames this as "cybersecurity as code," machine-speed threats require machine-speed defenses.

• Bitsight capability: Security posture management provides continuous, evidence-based measurement of your security controls so teams can identify gaps before attackers do and demonstrate improvement over time to regulators and leadership.

Step 5: Extend your defense to your third-party ecosystem.

Frontier AI attacks do not stop at your perimeter. The NYDFS advisory specifically directs entities to map and coordinate with third-party service providers on downstream risk. The CCCS similarly notes that cyber resilience is a shared responsibility extending across the entire supply chain. A compromise at a vendor with access to your systems or data is operationally equivalent to a direct breach, and frontier AI is accelerating the discovery of third-party weaknesses.

• Bitsight capability: [Third-party risk management software enables continuous monitoring of your vendor ecosystem's security posture, so you can identify supplier-side exposures and respond before they become your incident.

Best Practices and Expert Tips for Frontier AI Defense

The five-step framework above provides structure. The following expert practices help operationalize it within real security programs.

• Compress your vulnerability management cycle to match attacker speed. Traditional monthly or quarterly patching cycles are no longer adequate for critical vulnerabilities. Bitsight Threat Intelligence data shows that exploitation of high-severity vulnerabilities can begin within hours of public disclosure when AI-assisted exploit adaptation is in play. Prioritize a continuous patching cadence for internet-facing systems and critical infrastructure, and use exploitation likelihood scores rather than static severity ratings to sequence remediation work.
• Treat your AI deployment layer as attack surface, not just a tool. The 360% increase in exposed AI-related tooling observed by Bitsight TI in 2025 reflects how rapidly organizations have deployed AI services without including them in standard attack surface inventories. Langflow instances, model APIs, AI coding assistants, and MCP-connected tool chains all need to be tracked, authenticated, and monitored as part of your external attack surface management program.
• Build underground AI signal into your threat intelligence program. Jailbreak toolkits and AI-assisted exploit workflows are actively traded and refined across underground communities. Bitsight TI observed that private jailbreak prompts shared via direct message, invite-only channels, or paywalled content are treated by threat actors as more operationally valuable than public prompts, meaning the most dangerous tradecraft is not visible in open forums. Effective threat intelligence programs need visibility into both public and restricted underground sources.
• Validate AI-generated code before it reaches production. The NYDFS advisory specifically calls out the need for additional testing and human oversight of AI-generated code before production deployment. Frontier AI tools make it faster to write code, and also faster to introduce vulnerabilities or, in adversarial scenarios, to embed malicious logic. Treat AI-generated code as untrusted until it has passed the same review gates as human-authored code.
• Extend third-party risk monitoring to cover your suppliers' AI exposure. The CCCS and NYDFS both emphasize that supply chain coordination is essential in a frontier AI threat environment. Continuously monitor the security posture of vendors and service providers, with specific attention to their AI-related attack surface. A vendor running an unpatched, internet-facing AI orchestration service is a risk to your organization, not just to theirs.
• Document your frontier AI risk decisions for regulatory review. Regulatory bodies including NYDFS have signaled that frontier AI guidance carries supervisory and examination weight, even where it is not strictly binding. Security leaders should document the risk assessment process, the controls they have adopted in response to frontier AI advisories, and the rationale for any measures they have chosen not to implement. This documentation is also useful for board-level reporting and cyber insurance underwriting.

Advantages and Benefits of Security Platforms Designed for Frontier AI Defense

Security platforms built to address machine-speed threats deliver measurable operational improvements over conventional, point-in-time approaches. The following benefits reflect what organizations achieve when they operationalize the five-step framework with purpose-built intelligence capabilities.

• Real-time attack surface awareness: Continuous external attack surface management eliminates the blind spots that frontier AI attackers exploit during reconnaissance. Organizations know their exposure before attackers map it.
• Exploitation-led vulnerability prioritization: Replacing CVSS-only scoring with exploitation likelihood intelligence allows security teams to focus remediation resources where the actual risk is highest, reducing the total number of vulnerabilities that need urgent attention and ensuring that the most dangerous ones are addressed first.
• Underground intelligence at scale: Processing 540 billion cyber events and 7 million intelligence items daily across more than 1,000 underground sources gives security teams the adversarial context they need to anticipate frontier AI-enabled attack techniques before they are deployed at scale against their organization.
• Continuous posture measurement for regulatory readiness: Ongoing security posture measurement provides the evidence base needed to respond to regulatory expectations from bodies like NYDFS, NCSC, ENISA, and the CCCS, all of which have emphasized that organizations must raise their security baselines in response to frontier AI.
• Supply chain risk visibility: Continuous third-party monitoring extends frontier AI defenses to the vendor ecosystem, where compromises increasingly originate. Organizations can identify and address supplier-side exposures before they become the entry point for an AI-accelerated attack.
• Faster mean time to detect and respond: When intelligence, exposure data, and posture measurement are unified on a single platform, security teams spend less time aggregating context and more time acting on it, a critical advantage when exploitation windows are measured in minutes rather than days.

How Bitsight Helps Security Teams Defend Against Frontier AI

Bitsight is the cyber risk intelligence platform that connects the five steps of frontier AI defense into a unified, continuous workflow. Where many organizations operate disconnected point solutions, a scanner here, a risk register there, Bitsight integrates attack surface intelligence, exploitation-prioritized vulnerability data, underground threat intelligence, posture measurement, and third-party risk monitoring on a single platform built for the speed that frontier AI demands.

Bitsight Threat Intelligence has been tracking the evolution of AI abuse since before it became a mainstream concern. The Bitsight State of the Underground 2026 and the From Jailbreaks to Agentic Attacks: The Evolution of AI Abuse report document the specific tradecraft, tooling, and campaigns that security teams need to understand to defend effectively, from the operationalization of jailbreak workflows to agentic ransomware campaigns and MCP-based compromise paths.

For CISOs preparing for board-level conversations about frontier AI readiness, Bitsight's security ratings provide an externally verifiable, continuously updated signal of security performance that regulators, insurers, and executive leadership can act on. For security operations teams, Bitsight's external attack surface management and cyber threat intelligence capabilities provide the continuous visibility and adversarial context that machine-speed threats require.

As Emma Stevens, Senior Threat Intelligence Advisor at Bitsight, observed in the State of the Underground 2026: "The threat landscape in 2025 became more distributed, more adaptive, and in several areas more mature." Frontier AI is the primary driver of that maturation, and responding to it requires intelligence and platform capabilities that keep pace.

Key Takeaways and Next Steps

Frontier AI security is not a future concern. Regulatory bodies on multiple continents, the NYDFS, NCSC, CCCS, ENISA, and UK AISI, have all issued formal guidance in 2026 acknowledging that the threat is operational today. The five-step framework covered in this guide, inventory your attack surface, prioritize by exploitation likelihood, monitor underground AI activity, measure your posture continuously, and extend your defense to third parties, provides a practical foundation for any security program.

The evidence from Bitsight Threat Intelligence is clear: attackers are integrating frontier AI into their workflows at speed, underground discussion of AI models and jailbreak techniques is operating at massive scale, exposed AI-related infrastructure has grown dramatically, and agentic attacks are graduating from proof of concept to active campaigns. Security teams that treat frontier AI as a future risk rather than a present one are already operating with a gap.

For a detailed breakdown of the tools and platforms evaluated against frontier AI defense capabilities, read this guide. For a CISO-focused readiness roadmap covering governance, regulatory alignment, and board reporting, see our CISO AI Readiness Guide.

Read the full research behind this guide: the Bitsight State of the Underground 2026 and From Jailbreaks to Agentic Attacks: The Evolution of AI Abuse. Or view Bitsight pricing to explore how the platform fits your security program.

FAQs About Frontier AI Security

What is frontier AI security?

Frontier AI security is the discipline of defending organizations against attackers who use the most advanced and capable AI models to accelerate, automate, and scale offensive operations. It is distinct from AI governance or securing your own AI deployments: the focus is on the external threat posed by adversaries wielding frontier models. Bitsight provides research, threat intelligence, and platform capabilities specifically designed to help security teams understand and respond to this evolving threat category.

Why do CISOs need frontier AI security tools in 2026?

Frontier AI has materially changed the speed and scale at which attackers can discover and exploit vulnerabilities. Regulatory bodies including NYDFS, NCSC, ENISA, and the Canadian Centre for Cyber Security have all issued formal guidance in 2026 directing organizations to prepare. Bitsight Threat Intelligence observed that exposed AI-related tooling grew 360% in 2025 to more than one million exposed services, an attack surface that is growing faster than most organizations have inventoried it. CISOs need tools that provide continuous visibility, exploitation-led prioritization, and underground intelligence at machine speed.

What is a Mythos-class attack and how does it differ from conventional cyberattacks?

A Mythos-class attack uses a frontier AI model to execute multi-step offensive operations with minimal human involvement. The UK AI Security Institute's evaluation of Claude Mythos Preview found that it could execute multi-stage attacks on vulnerable networks and autonomously discover and exploit vulnerabilities, tasks that would take human professionals days. Conventional attacks depend on human operators moving through each phase manually. Mythos-class attacks compress the entire kill chain, from reconnaissance to exploitation, into a timeline that traditional detection and response programs were not designed to address.

Which security platforms protect against frontier AI threats?

Effective frontier AI defense requires platforms that provide continuous external attack surface management, exploitation-prioritized vulnerability intelligence, underground threat intelligence, security posture measurement, and third-party risk monitoring. Vendors including Bitsight, as well as others such as Akamai, SentinelOne, and Wiz, offer capabilities relevant to parts of this framework. Bitsight's cyber risk intelligence platform integrates all five capabilities, backed by intelligence drawn from more than 540 billion cyber events and over 1,000 underground sources. For a detailed comparison, see this guide. 

How should CISOs prepare for frontier AI threats?

CISOs should act across five dimensions: continuously inventory and monitor the external attack surface including AI infrastructure, accelerate vulnerability remediation based on exploitation likelihood rather than static scores, integrate underground AI abuse signals into threat intelligence programs, continuously measure and evidence security posture for regulatory reporting, and extend frontier AI risk management to their third-party vendor ecosystem. Bitsight's research reports, Bitsight State of the Underground 2026 and From Jailbreaks to Agentic Attacks: The Evolution of AI Abuse, provide the adversarial context needed to inform each of these steps.

What is MCP tool poisoning and why does it matter for frontier AI defense?

Model Context Protocol is a protocol that allows AI agents to connect with external tools, files, APIs, system environments, and internal resources. Bitsight Threat Intelligence found that MCP servers can expose files, environment variables, internal APIs, repositories, and shell-backed tools to AI agents. Tool poisoning occurs when an adversary manipulates MCP-connected tools or prompts to redirect an AI agent's actions toward malicious outcomes. MCP tool poisoning matters because it creates a compromise path that does not require conventional malware and may not trigger traditional endpoint or network detection signatures.

How does Bitsight's DVE Intelligence help with frontier AI defense?

Bitsight's DVE (Distributed Vulnerability Exploitation) Intelligence scores vulnerabilities based on real-world exploitation activity observed across the threat landscape, not solely on theoretical severity metrics. In a frontier AI environment, where exploitation windows can collapse from weeks to hours, the ability to identify which vulnerabilities are actively being targeted, and prioritize those above all others, is operationally critical. The Bitsight State of the Underground 2026 found that all 10 highest-impact 2025 Known Exploited Vulnerabilities carried both Bitsight DVE and CVSS scores above 9, validating that DVE Intelligence identifies the vulnerabilities that matter most before they become breaches.

*Last updated: October 2, 2026. This page covers the frontier AI security threat landscape based on research and regulatory guidance current as of the publication date. The field is evolving rapidly; readers are encouraged to monitor updates from NCSC, ENISA, CCCS, NYDFS, and Bitsight Threat Intelligence for the latest developments.*