Defending against AI-Equipped Attackers: A 3-Layer Approach

How can your business defend its attack surface from adversaries who can use public AI large language models (LLMs) at relatively low cost to autonomously discover zero-day vulnerabilities across your critical business and IT systems faster than you can patch them or test for them?

A successful defense requires a 3-layer approach based on zero-trust principles:

1. Reduce reach—Limit what attackers can access.
2. Harden identity—Make credentials and access insufficient to breach an application or system.
3. Survive the attack—Minimize the damage successful attacks can cause.

Read on to get the details and a 90-day rollout plan.

Why are AI-enabled threats breaking traditional cybersecurity?

AI-enabled threats break traditional cybersecurity by accelerating attack timescales from weeks to hours or minutes and operating autonomously at machine speed. As a result:

• AI LLMs like Anthropic Claude Mythos can now find zero-day vulnerabilities and create exploits much faster than humans can patch them. 
• Traditional safeguards like signature-based anti-malware tools depend on known patterns. These are useless against novel attacks and sophisticated “fileless” intrusions or “living off the land” tactics. 
• AI agents can dynamically probe the target environment in real-time to move laterally around defenses.
• The threat landscape now changes continuously, reducing the value of periodic vulnerability scans and penetration testing.

In short, the longstanding view that vulnerabilities in production software are scarce and slow to surface no longer holds true. Organizations can no longer hope to harden their perimeters and configurations enough to dependably protect sensitive data with conventional, human-speed solutions.

Instead, a 3-layer defensive strategy based on zero-trust principles, least privilege access, best-practice cybersecurity, and “AI on your side” is required. The sections that follow explain the 3 layers and what each should include.

Layer 1: Reduce reach.

Reducing reach means limiting what attackers can access in your environment. This requires you to map all your assets, segment systems and services, take basic steps to shrink your attack surface, and limit the blast radius when (not “if”) compromises occur.

Critical steps to reduce an attack’s potential reach include:

Asset inventory and visibility

• Develop and maintain real-time asset and software bill of materials (SBOM) inventories. You can’t protect what you don’t know exists.
• Dynamically map all your in-house and third-party applications against their associated physical/data center locations, overarching business processes, or network segments.
• Keep your asset inventory continuously updated.

Segmentation of networks, systems, and services

• “Never trust, always verify.” Make all users and digital entities confirm their identity (e.g., with multifactor authentication steps) before being given access to sensitive data.  
• Create strong micro-segmentation between and around critical systems. 
• Enforce least-privilege access between services. Give users and devices only the bare minimum access level required to perform their current specific task.

Attack surface reduction

• Decommission legacy systems and disused services.
• Eliminate all unnecessary internet exposure.
• Consolidate redundant tools and reduce IT/vendor complexity.

Contain the blast radius

• Assume that hackers will inevitably gain initial access to your systems. If it hasn’t happened already, it will.
• Mitigate digital risks so that attacks can be quickly contained, not potentially catastrophic.
• Ensure that your most critical data (e.g., customer data, intellectual property, financial records) is protected from the rest of the network by controlled paths enforced through network segmentation and micro-perimeters. This blocks an attacker from moving laterally around your network, contains the spread of ransomware/extortionware, and triggers alerts in response to suspicious actions.  

Layer 2: Harden identity.

Hardening identity means adding controls that prevent an attacker from gaining unauthorized entry to a system or application even if they have account credentials and network access. 

Key steps to safeguard critical systems from credential-based attacks include:

Phishing-resistant MFA

• Use FIDO2 security keys or passkeys instead of passwords for all privileged accounts. This reduces the risk to sensitive data from phishing attacks and credential theft. 
• Retire SMS, push notifications, and one-time codes in high-risk access scenarios. These traditional protections are vulnerable to phishing and AI-powered attacks. 
• Move to hardware-bound, cryptographically verified credentials for systems where the stakes are highest. Pairing credentials with a specific device/application means there is no code for attackers to intercept or relay. 

Identity-centric incident detection

• Deploy behavioral anomaly detection for improved cybersecurity and fraud prevention, especially against AI-powered actors. 
• Combine single-tool alerts for cross-system correlation analysis to reveal multi-step attack paths and track machine identities.
• Leverage identity threat detection and response (ITDR) as a frontline control to protect from credential-based, AI-driven attacks. Identity is now the primary attack surface, and ITDR helps spot anomalous behavior after authentication. 

Just-in-time privileges

• Eliminate standing, permanent, or “always on” privileged/administrative access associated with a user or system account wherever possible. Permanent privileged access increases the risk of catastrophic incidents by yielding immediate full control of the asset with no further effort required.
• Replace standing admin access with temporary permissions granted only when necessary. Elevation on demand requires users or systems to request short-term access for special tasks using an automated workflow, while time-bound permissions have a predefined “expiration date” (usually within 2 hours). An associated immutable audit trail is ideal for compliance and threat analysis. 
• Choose a privileged access management (PAM) tool that fits your current workflows, business structure, and IT stack. This reduces resistance, improves productivity, and lowers IT effort and cost. Avoid forcing staff to retrofit operational processes to comply with rigid software rules.

Service and machine identity protections

• Discover, inventory, and track non-human identities (e.g., machine accounts, service accounts, API keys, tokens) to manage them throughout their lifecycle from creation through deletion. This improves access control and shrinks the attack surface by eliminating unused identities.
• Rotate, scope, and monitor non-human identities to prevent their misuse, especially by malicious AI agents. 
• In the transition from human-centric to machine-centric identity management, treating autonomous AI agents as fully accountable digital entities with their own, unique identifiers is the next governance frontier. Letting AI agents share human logins or giving them static service credentials confuses accountability, invites privilege creep, and gives AI-powered attackers more chances to move laterally within the network.

Layer 3: Survive the attack.

To survive an attack, you need to plan to recover from what you cannot prevent. Incidents are inevitable in today’s threat landscape—but organizations can greatly reduce the risk of overwhelming financial, operational, and/or reputational damage by containing the impact.

Essential steps to minimize impacts when a cyber-attack penetrates your defenses include:

Behavior-based detection

• Deploy XDR/MDR tuned to user behavior, not known malware signatures. Many AI-driven attacks are novel and proceed autonomously rather than in preprogrammed steps.
• Deploy network behavior analytics (NBA) to detect attackers’ lateral movement. NDA monitors network traffic and uses machine learning to alert on potential threats that stand out against the normal activity baseline.
• Use detection tools that do not rely on prior knowledge of the exploit, as malicious AI agents often use new and innovative attack paths. 

Immutable, tested recovery

• Isolate backups from the production network and the universe of production identities. This prevents a threat actor that breaches your operational environment from using what it knows to damage your recovery data or propagate extortionware encryption into your backups through real-time replication or synchronization. 
• Test your recovery process end-to-end, at scale. This is what separates a proven process from an unproven plan. Full-scale testing is essential to uncover hidden dependencies, validate recovery time and recovery point objectives (RTO/RPO), train staff, and shake out surprise failure points. 
• Get management to sign off on your RTO and RPO. Present options that connect IT solutions and capabilities to business cost and risk. You will first need to analyze the financial and reputational impacts of downtime, along with identifying critical versus less essential systems.  

Rehearsed response

• Develop incident response playbooks for the most likely threat scenarios your business faces. This helps cut response times, eliminate guesswork, prevent missteps, and minimize damage. 
• Conduct tabletop and live-fire incident response exercises on a regular basis to improve performance, identify gaps, test changes, and validate technical capabilities (e.g., RPO and RTO). 
• Proactively agree on decision rights and escalation paths for your incident response plan. Otherwise, you might be debating who has authority or requires notification in the midst of an outage, leading to costly delays and post-incident finger-pointing. 

AI on your side

• Run AI-assisted vulnerability scans against your own environment. Don’t wait for hackers to discover the vulnerabilities for you. 
• Use AI to filter, prioritize, and enrich security alerts in real-time by incorporating threat intelligence data. 
• Use the same autonomous AI capabilities to protect your environment as the adversary is using against you. This is imperative because traditional, human-speed security tools cannot contend with the pace, scale, and autonomous response capability of today’s AI-driven attacks. 

A 90-day priority list to start building a 3-layer defense

Building a robust 3-layer defense against agentic AI malware takes significant organizational commitment, technical expertise, and resources. But the cost of a data breach is likely to be much, much greater.   

Here are the 8 most important steps to prioritize in the first 90 days of your timeline:

1. Move from an annual or periodic asset and SBOM list to a real-time inventory.
2. Identify and decommission your top 10 unused or unnecessary services.
3. Protect all privileged accounts with phishing-resistant MFA.
4. Eliminate standing/static access privileges wherever they are not absolutely required.
5. Test your backup process end-to-end against a realistic ransomware scenario. 
6. Run a tabletop incident response exercise against realistic agentic AI malware.
7. Run AI-assisted vulnerability discovery against your own proprietary code base.
8. Brief your board, senior management, and other key stakeholders on this massive shift in the threat landscape and what it means for current and future cybersecurity investments.

What’s next?

Concerned about AI-powered threats and how to defend against them? Partnering with CBIZ Pivot Point Security gives you a trusted advisor to help establish the foundational controls your business needs for resilience and growth.