iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
AI is changing cybersecurity on both sides: it can help defenders find weaknesses and respond to threats, while helping attackers speed up reconnaissance, vulnerability discovery, exploit development, and social engineering. The near-term balance may favor attackers, according to a 2026 National Academies expert consultation, but that is an assessment—not a universal or quantified result. The practical challenge is to secure AI systems themselves while keeping conventional defenses, oversight, and remediation able to keep pace.
How AI is changing cybersecurity
Generative AI, reasoning systems, and agents can assist with tasks across the cyberattack lifecycle. The National Academies’ 2026 rapid expert consultation identifies reconnaissance, vulnerability discovery, exploit development, and social engineering as areas that may become faster, cheaper, or easier to scale. Defenders can also use AI to analyze information and identify weaknesses for remediation.
The difference is not simply that one side has AI and the other does not. Defensive work such as validating alerts, coordinating responses, maintaining complex software, and carrying out remediation involves many systems and teams. Automating part of a task does not automatically compress the coordination and judgment needed to finish it safely. That asymmetry underlies the consultation’s concern that attackers could gain a near-term advantage.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsIt is not established that AI has produced a lasting advantage for attackers across sectors, or that a particular capability demonstration improves security outcomes in a real organization. The National Academies notes that widely accepted frameworks for evaluating AI-enabled cyber capabilities and tracking their evolution are lacking. Results depend on the systems being used, the surrounding defenses, and how they behave under adversarial conditions.
#1 Best Overall
Two different problems: attacks on AI and AI used in attacks
Organizations need to distinguish attacks that target an AI system from attacks in which a person uses AI to target conventional systems or people. The two can overlap, but they call for different controls. NIST describes AI as both a potential aid to defenders and a technology that can enhance attacks on information technology and operational technology.
| Problem | What may be targeted or affected | Why it matters |
|---|---|---|
| Attacks on AI systems | Inputs, model data and weights, serving infrastructure, user data, outputs, and connected supply-chain components | An attacker may manipulate, steal, expose, or disrupt parts of the AI service. NIST emphasizes that confidentiality, integrity, and availability concerns apply to AI as they do to other software and data systems, alongside AI-specific attack surfaces. |
| AI used to attack other targets | Conventional computer systems, people, sensitive information, and—in some contexts—physical capabilities | AI may assist with tasks such as reconnaissance or social engineering without the AI system itself being the target. |
NIST’s 2025 taxonomy gives useful names to adversarial machine-learning techniques without implying that each is prevalent in every deployment. For predictive AI, it includes evasion, poisoning, and privacy attacks. For generative AI, it includes those categories plus misuse attacks. The taxonomy also discusses possible mitigations and their limitations; a named attack category is a way to reason about risk, not proof that a particular system is vulnerable.
Why agents change the control problem
An AI agent may have an identity, access to data, connections to tools, and permission to take actions. That creates risks beyond whether the model produces a safe or accurate answer. A compromised or misdirected agent with broad permissions can affect systems and information beyond the conversation in which it operates.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →A joint cybersecurity information sheet announced by the NSA and partner agencies on April 30, 2026 identifies privilege, design and configuration, behavior, structural, and accountability risks. Microsoft’s 2026 Digital Defense Report groups agent risks as prompt and intent manipulation, sensitive-data exposure, identity and privilege compromise, excessive agency, and operational integrity. These frameworks describe related concerns from different perspectives:
Rank #3
- Identity and privilege: an agent’s credentials or permissions may be compromised or broader than its task requires.
- Tools and connected components: integrations increase the number of systems an agent can reach and the ways it can affect them.
- Instructions, memory, and data: manipulated prompts or poorly scoped access can steer behavior or expose sensitive information.
- Actions and accountability: an agent may take consequential actions that are difficult to review if behavior is opaque or audit trails are weak.
Controls in Microsoft’s framework include prompt inspection, scoped memory and credentials, verifiable agent identity, least privilege, tool allow-lists, runtime action gates, anomaly detection, and protection for configurations and supply chains. These are examples of layered safeguards, not one-to-one guarantees against each risk. The joint agency guidance calls for incremental deployment, continuous assessment against evolving threat models, clear governance and accountability, rigorous monitoring, and human oversight.
What organizations can do now
1. Keep core security controls in place
AI changes the threat environment; it does not make established cybersecurity practices optional. Microsoft’s 2026 guidance recommends rapid patching, access control, resilience, secure-by-design development, Zero Trust, multifactor authentication (MFA), least privilege, and ongoing security training. These controls reduce exposure but are not guarantees against AI-enabled attacks. For identity protection, MFA can include a FIDO2-compatible hardware security key; the cited sources do not compare or endorse particular products.
2. Map every AI service’s reach
For each AI service or agent, identify the data it can read, the identities and credentials it uses, the tools and connected systems it can reach, and the actions it can perform. Remove access that is not necessary for its task. Put validation or human approval in front of actions with significant consequences, and ensure activity can be monitored and audited. These steps apply the risk areas and controls described in Microsoft’s report and the joint agency guidance.
Free tools Windows power users keep installed
One-click scans. No signup required.
3. Include AI in system and supply-chain security
Protect the model, its data, serving infrastructure, configuration, outputs, and connected components as parts of the system—not as isolated AI features. NIST describes work on implementation-oriented control overlays for large language models, predictive AI, agents, and AI developers. Those overlays are under development; they should not be described as completed final standards.
Best Value
4. Make vulnerability handling keep pace
Faster discovery only helps if an organization can validate findings, assess exploitability and real-world risk, assign an owner, and remediate. Establish who triages a finding, who can authorize a response, and how remediation capacity is allocated. An expanding queue of unvalidated findings is not the same as reduced risk.
5. Evaluate outcomes in the real deployment
Test how the system and its surrounding controls perform in the context where they will be used, including under manipulation and failure conditions. Track whether findings lead to verified fixes, whether agents stay within their permitted scope, and whether incidents can be detected and investigated. A capability demo alone cannot show that an organization is more secure.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the reported adoption figures do—and don’t—show
Microsoft’s 2026 Digital Defense Report includes several figures about AI agents and security work. They describe reported experimentation, plans, projections, or a vendor-associated result; they are not independent measures of AI’s net effect on cybersecurity.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →| Figure | What Microsoft reports | Qualification |
|---|---|---|
| 88% | Enterprises are already experimenting with AI agents. | Microsoft’s report passage does not identify the survey methodology, so this should be read as a Microsoft-reported figure, not an independently verified global estimate. |
| 82% | Leaders plan broader agent rollouts within 12 to 18 months. | This is a reported plan and timeframe, not evidence that the rollouts occurred or that they were secure. |
| Roughly 1.3 billion | AI agents in production by 2028. | Microsoft reports this as an industry projection; the passage reviewed does not name the originating organization. |
| 60 to 70% faster | Threat summarization among organizations using Microsoft Security Copilot. | This is a Microsoft-reported, product-associated result—not an independent measure of AI’s effect on security operations generally. |
These figures indicate activity and expectations, not whether deployments are resilient or whether attackers gain more than defenders. The National Academies’ consultation does not provide a broadly generalizable numeric estimate of AI’s net effect on cybersecurity outcomes.
Quick Recap
Sources and scope
- NIST, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, announced March 24, 2025.
- NIST, AI Research – Security and Resilience, accessed October 7, 2026; describes active research and proposed control overlays.
- National Academies of Sciences, Engineering, and Medicine, Implications of Recent Advancements in Artificial Intelligence for Cybersecurity, 2026 rapid expert consultation.
- Microsoft, Microsoft Digital Defense Report 2026; its report description says threat data covers July 1, 2025 through June 30, 2026. Findings attributed to Microsoft are vendor-reported.
- NSA, announcement of joint guidance on careful adoption of agentic AI services, April 30, 2026.
- Amy Hogan-Burney, Microsoft Corporate Vice President for Customer Security & Trust, From capability to responsibility, May 1, 2026.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
