AI is changing cyber operations first by compressing work, not by making every attack autonomous. Reconnaissance, vulnerability discovery, exploit development and social engineering can be turned into faster, lower-barrier workflows. Defensive work also benefits from AI, but it still has to connect monitoring, validation, response authority, patching, software maintenance and coordination across real organizations.
The National Academies’ June 2026 rapid expert consultation therefore describes a likely near-term advantage for attackers without concluding that defenders are destined to lose. The decisive question is whether organizations can turn AI-assisted detection and remediation into a continuous, coordinated defensive process.
As an Amazon Associate I earn from qualifying purchases.
How is AI changing the relative pace of cyber offense and defense?
The asymmetry is about the shape of the work. An attacker may need to complete only the next useful step: find an exposed service, identify a likely weakness, develop a working technique or persuade a target. AI can compress those steps and lower the expertise needed to attempt them. A defender must keep an entire environment safe, determine whether an alert is real, contain activity, repair the underlying weakness and verify that the repair worked.
The National Academies’ June 2026 rapid expert consultation says recent AI advances are likely to advantage attackers in the near term because key attack stages can be compressed. It also stresses that defensive processes can benefit from AI, but are harder to compress because they cross systems, teams and operational decisions.
#1 Best Overall
| Part of the operation | How AI can change the pace | What still limits the result |
|---|---|---|
| Reconnaissance | Organizes and analyzes information quickly, helping identify promising targets or paths. | The output still needs context about the target environment and authorization. |
| Vulnerability discovery | Assists with finding weaknesses and prioritizing areas for investigation. | A finding is not a confirmed exploitable condition, and remediation remains a separate task. |
| Exploit development | Lowers the barrier to producing or adapting techniques against a weakness. | Reliability depends on the specific software, configuration and defenses in place. |
| Social engineering | Can help generate convincing, targeted messages and variations. | Success depends on the target, delivery channel and defensive controls. |
| Monitoring and detection | Can process large volumes of telemetry and help analysts investigate alerts. | False positives, missing data and unfamiliar system context still require human and organizational judgment. |
| Response and repair | Can support triage, vulnerability prioritization and parts of investigation. | Containment authority, change control, testing, patch deployment and verification remain operational responsibilities. |
This is a workflow comparison, not a measured head-to-head race. The available expert assessment does not provide a universal attacker-versus-defender speed ratio or prove that an AI-assisted attack will succeed. It identifies where compression is easier and where defensive work remains integrated and consequential.
What does a “near-term attacker advantage” actually mean?
It means that attackers may reach a useful capability with fewer people, less time or less specialized knowledge. A criminal group can use an AI system to help with a stage of an operation without building a fully autonomous agent. The advantage can appear as more attempts, faster iteration or a broader pool of targets rather than as a single machine running an entire intrusion independently.
That distinction matters. The National Academies does not establish that AI attacks are universally autonomous, unstoppable or always superior to human-led operations. AI output remains dependent on prompts, data, access, tooling and the surrounding infrastructure. A generated suggestion can be wrong, unsafe or irrelevant to the target. Human operators may still select objectives, obtain access, adapt to defenses and decide when to act.
Defenders face a different risk: even if each AI-generated action is imperfect, lower costs can let an adversary try more paths. Security teams therefore need controls that reduce exposure and shorten the time between detection, containment and repair, rather than relying on the hope that an attacker’s model will make a mistake.
Why is it so difficult to assess AI cyber capabilities?
Capability measurement is itself a major roadblock. The National Academies reports that there is no widely accepted framework for evaluating AI-enabled offensive or defensive cyber capability over time. Organizations can measure a model on an individual task, but that score does not directly establish improved security outcomes in a live environment.
Task benchmarks are not operational resilience
A benchmark might test whether a system can identify a class of vulnerability, classify an alert or generate code. A real security outcome asks different questions: Did the system find the weakness in this organization’s software? Did experts confirm it? Was it repaired before exploitation? Did the change avoid breaking a critical service? Did monitoring detect a related attempt and did responders contain it?
Those questions involve data quality, system architecture, staffing, permissions, change windows and incident procedures. A high score on one task can coexist with weak telemetry, unpatched dependencies or a response team that cannot authorize emergency changes.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteGenerative behavior has no automatic guarantee
Generative systems are probabilistic. Their behavior can vary with inputs, context and model updates, and they can produce plausible but incorrect instructions. Security leaders cannot treat a model’s apparent competence as a guarantee of repeatable behavior, safe handling of sensitive data or resistance to manipulation.
Evaluation must therefore include the deployment context: the model’s tools, data sources, permissions, logging, human review and failure handling. It should test both useful performance and unsafe behavior, including what happens when the model receives misleading or adversarial input.
Compare capability claims with security outcomes
| Question | Capability evidence | Outcome evidence needed |
|---|---|---|
| Can the system find weaknesses? | Accuracy or coverage on a defined task set. | Confirmed findings in the organization’s code and infrastructure, followed by timely, verified remediation. |
| Can it analyze alerts? | Classification or summarization performance. | Faster, more accurate triage without missed incidents or unmanageable false positives. |
| Can it assist response? | Quality of proposed actions or playbooks. | Containment that is authorized, safe, logged and effective under realistic incident conditions. |
| Can it resist manipulation? | Results in a controlled test. | Stable behavior when exposed to adversarial inputs, untrusted data and changing system conditions. |
Without this distinction, organizations can mistake a compelling demonstration for evidence that a deployment improves resilience.
Rank #3
What is the clearest official evidence of AI-assisted malicious activity?
The opening statement for the U.S. intelligence community’s 2026 Annual Threat Assessment reports that, in August 2025, cyber actors used an AI tool in a data-extortion operation targeting international government, health care, public health emergency services and religious institutions. The account appears in the Office of the Director of National Intelligence’s 2026 opening statement.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →This is an official example of AI use in a malicious operation. It does not show that the operation was fully autonomous, that the tool performed every stage, or that AI was the only reason it succeeded. The evidence supports a narrower conclusion: threat actors are using AI as part of real activity, including data extortion, against high-consequence sectors.
That qualification should shape incident analysis. Investigators should record which model or service was used, what data it accessed, which actions were automated, where a human approved or changed the output, and which conventional infrastructure enabled the operation. “AI-assisted” is more informative than treating every event as an autonomous attack.
What roadblocks keep defenders from catching up?
Defense has to cross systems and teams
An attacker can focus on a narrow objective. A defender must connect endpoint, identity, network, application and cloud signals; maintain asset inventories; validate alerts; coordinate legal, communications and business decisions; and preserve evidence. The relevant data may sit with different teams or suppliers, each with separate access rules and response authority.
AI can help analyze those inputs, but it cannot create missing logs, grant permission to isolate a production system or make a patch safe for an untested dependency. Integration work is an organizational requirement, not a feature that appears automatically when a model is deployed.
Validation is slower than generation
Attackers can discard a failed generated idea and try another. Defenders must validate a suspected weakness before changing a critical system, and they must test whether a fix closes the actual path without introducing a new outage. This makes false positives expensive and explains why vulnerability discovery alone does not equal risk reduction.
AI-supported discovery and prioritization are valuable when they help experts focus on exploitable, high-impact weaknesses. The repair process still needs ownership, testing, deployment and confirmation.
Coordination determines response speed
Even an accurate detection is useful only if someone can act. Organizations need defined escalation paths, emergency change procedures, contact information for shared infrastructure and agreements covering suppliers or partners. Where software and response responsibilities are distributed, catching up requires sustained coordination rather than a one-time tool purchase.
Dual-use techniques can help both sides
Methods for finding weaknesses, generating code or analyzing behavior are dual use. The same capability can support secure development or help an attacker locate an exploitable path. Access controls, audit logs, segmented environments and review gates are needed to keep useful security work from becoming an uncontrolled source of offensive capability.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →AI creates another attack surface
AI systems must be protected as operational assets. The National Institute of Standards and Technology (NIST) notes that AI security overlaps with conventional software and data security while adding concerns such as evasion, model extraction, membership inference and availability attacks.
Best Value
In practical terms, a defender must protect more than the model’s output. Data pipelines, model weights, configuration, credentials, tool connections, access policies, update processes and service availability can all affect security. A compromised model or manipulated input may produce an unsafe recommendation even when the surrounding application appears intact.
NIST is developing control overlays for large language models, predictive AI, agents and AI developers, along with a test platform for evaluating vulnerabilities and defenses. Its Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations provides a common vocabulary for discussing adversarial techniques and mitigations. These resources do not remove the need for local testing; they help organizations describe and evaluate the risks consistently.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What can defenders do to respond to AI-enabled threats?
The following actions address the operational gaps identified by the expert and government sources. They are risk-reduction measures, not a guarantee against compromise.
- Raise baseline security and software quality. Maintain accurate asset and dependency inventories, apply secure development practices, reduce unnecessary exposure and fix high-impact weaknesses. Strong fundamentals limit the number of easy paths an attacker can exploit, regardless of whether AI is involved.
- Use AI to accelerate discovery and prioritization, not to skip verification. Apply AI-assisted analysis to code, configurations, telemetry and vulnerability queues where it can help experts identify likely impact and urgency. Require confirmation, ownership and a documented repair path for each material finding.
- Shorten the detection-to-response loop. Ensure important systems generate usable logs, route alerts to accountable teams and support tested containment procedures. Measure whether analysts can validate an alert and whether responders can act within the organization’s actual authority and change process.
- Share threat intelligence quickly. Exchange indicators, techniques, affected software details and lessons from incidents with relevant partners. Shared information can reduce duplicated investigation and help organizations recognize AI-assisted activity without assuming that every event has the same technical signature.
- Secure AI systems as production infrastructure. Control access to data, model weights, prompts, tools and administrative functions. Log model and tool activity, protect configuration and update paths, test availability, and plan for model or data compromise. Include evasion, extraction, inference and adversarial-input scenarios in security reviews.
- Evaluate in the environment where the system will operate. Test realistic data, permissions, integrations, failure modes and human handoffs. Compare benchmark results with operational measures such as confirmed findings, response time, false-positive burden, safe change completion and recovery performance.
- Invest across organizational boundaries. Align security, engineering, operations, legal, procurement and leadership on who owns a system, who can authorize an emergency action and how outside providers participate. Continuous defense depends on these relationships as much as on model quality.
How should organizations avoid overestimating AI on either side?
- Do not equate faster generation with successful exploitation. A generated exploit or message still has to work against a particular target and its controls.
- Do not equate a benchmark score with protection. Security improvement requires evidence from the deployed system and the surrounding process.
- Do not call tool use autonomy without evidence. Document the human decisions, permissions and infrastructure involved in an operation.
- Do not treat AI security as separate from ordinary security. Weak identity, access, software quality, logging or availability controls can undermine an AI deployment just as they undermine other systems.
- Do not automate irreversible actions without safeguards. Review, approval, rollback and verification are especially important when a model can change systems or invoke external tools.
Can AI help defenders catch up over time?
Yes, but only conditionally. The National Academies says continuous AI-driven defense in depth could shift the balance toward defenders if investment and coordination are sustained. That possibility depends on combining better software quality, rapid vulnerability remediation, strong monitoring, effective response and protected AI systems.
The outlook is therefore neither “AI guarantees attacker victory” nor “a defensive model will solve cybersecurity.” Attackers may gain an early advantage from compressed workflows, while defenders can build a countervailing advantage through scale, continuous coverage and institutional coordination. Whether that happens will be determined by implementation, measurement and governance—not by a benchmark score or a claim of autonomy.
Quick Recap
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.
