PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchiTechGuides is reader-supported.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
Business leaders should treat AI as part of the organization’s existing security environment—not as a separate problem that replaces conventional cybersecurity. Protect the data, software, hardware and accounts around AI systems with established controls, then add safeguards for model-specific threats and for any tools or actions an AI can access. A July 2026 survey found significant reported attacks and gaps in detection and response, but its results describe respondents, not every business.
What changes when AI enters the attack surface?
AI can expand both an organization’s capabilities and its exposure. A model may process confidential inputs, generate outputs used in business decisions, or connect to other software. An AI agent may also be able to take actions on a user’s behalf. Each connection creates security questions about what data and systems are reachable, what the model or agent is permitted to do, and how people can detect and contain misuse.
The basic objectives remain familiar: confidentiality, integrity and availability. NIST describes AI security as an extension of established software and information-system security concerns, with additional AI-specific attack types and attack surfaces. It also notes AI’s dual-use character: AI may help defenders address vulnerabilities and may strengthen the capabilities of attackers targeting information technology and operational technology environments. That is a reason to assess AI-related exposure, not evidence that every incident is AI-driven. NIST’s AI security and resilience overview outlines this framing.
Threats to include in AI-specific reviews
Keep conventional threats in scope, and consider how AI changes the attack path. NIST’s security overview identifies threats such as evasion, model extraction, membership inference and attacks on availability. The relevant risks depend on the system, the data it handles and its connections; a model used for internal analysis does not necessarily have the same exposure as one that can invoke business tools.
#1 Best Overall
Which familiar controls still matter?
AI does not remove the need for ordinary security fundamentals. Inventory the systems and data involved, restrict access, protect data, secure development, monitor activity, patch underlying software and respond to incidents. Apply these controls to the full AI environment: its applications, infrastructure, integrations and the data and configuration that support it.
Threat modeling should include the AI-specific assets and attack paths. NIST’s security framing covers training and output data as well as underlying software and hardware; NIST’s COSAiS project also considers components such as training and test data, model weights and configuration settings. For each system, ask who can access those assets, what changes could compromise their integrity, and what interruption would affect availability. NIST’s COSAiS project page describes the system types and components being considered in its evolving guidance.
Rank #2
Why assistants, predictive AI and agents need different controls
A useful review starts with the system’s actual boundary, not the label “AI.” NIST’s COSAiS work spans generative AI assistants and large language models, predictive AI, single-agent and multi-agent systems, and AI developers. These categories can expose different assets and paths to harm. The following comparison is an editorial way to apply those distinctions, not a formal NIST risk score.
| System type | Boundary to examine | Security question |
|---|---|---|
| LLM assistant | Prompts, connected data, outputs and any integrations | Can sensitive information enter or leave the system, and who can access its outputs? |
| Predictive AI | Input and test data, model, configuration and decisions that use its predictions | Could manipulated inputs or unauthorized changes undermine the prediction or downstream decision? |
| Single agent | Tools, accounts and data available to the agent | What actions can it take, and which actions require a person’s approval? |
| Multi-agent system | Interactions among agents and their shared tools or information | How are permissions, handoffs and actions controlled across the system? |
| AI development pipeline | Training and test data, model weights, configuration and supporting software | Who can change or access each component, and how are changes monitored? |
For assistants and agents, access and autonomy deserve particular attention. Inventory the tools and data the system can reach, grant only the access needed for its purpose, and require human review where an action could have material consequences. Keep logs and test whether the organization can recognize and investigate unexpected behavior. These controls should follow the system’s real capabilities: a chatbot that only drafts text and an agent that can change records have different action boundaries.
What the 2026 survey does—and does not—show
The SANS Institute’s July 2026 survey drew on responses from 536 security practitioners and 57 senior security leaders globally. Its figures are survey findings, not audited population statistics or proof that every organization faces the same incident pattern. The SANS 2026 AI in Cybersecurity report reports the following:
- 78% of organizations reported confirmed or suspected AI-enabled attacks — SANS Institute, 2026. “Confirmed or suspected” is the reported survey category; it should not be read as independent verification of every case.
- 95% of respondents believed threat actors were already using AI — SANS Institute, 2026. This measures belief, not incident verification.
- 78% of security teams actively used AI in cybersecurity in 2026, compared with 50% in 2025 — SANS Institute, 2026. The figures indicate reported adoption, not that every deployment is mature.
- 27% of practitioners called their AI deployment mature production — SANS Institute, 2026.
- 63% of practitioners reported significant AI shortcomings in threat detection and response — SANS Institute, 2026.
- 76% of security teams held a governance role for enterprise AI — SANS Institute, 2026.
- 73% of practitioners said AI changed their team’s training requirements in 2026 — SANS Institute, 2026.
- 61% of practitioners used AI in red-team work — SANS Institute, 2026.
The report also describes a 14-point perception gap about whether a formal AI risk management program exists: 50% of senior leaders versus 36% of practitioners said one existed. Alongside the reported governance role and low share calling deployments mature production, that difference is a useful prompt for leaders to check whether policy, ownership and operational practice agree within their own organization. It is not evidence that the same gap exists at every company.
How leaders can connect governance to operational security
Governance is useful when it clarifies who owns a system, what it is allowed to do and how the organization will respond when it behaves unexpectedly. Security teams may advise or oversee enterprise AI, but business owners, system developers and operators also need clear responsibilities. The goal is to make risk decisions and escalation paths usable in day-to-day work.
- Map AI use and ownership. Identify systems built internally, procured externally or embedded in workflows; name the business owner and the teams responsible for security and operation.
- Record assets, access and autonomy. For each system, document its data, model-related components, integrations, permissions and ability to act without approval.
- Set controls proportionate to exposure. Use established security controls as the baseline, then add safeguards for the AI-specific attack paths and consequential actions relevant to that system.
- Validate detection and response in context. Test model-assisted detection in the organization’s own environment. Define how staff investigate an alert, verify AI-generated findings and contain an incident without relying on an unvalidated output.
- Train the people who build, use and secure the system. Cover the system’s access boundaries, AI-specific risks, safe handling of data and escalation procedures. The SANS survey’s training finding supports treating capability as an active security concern, not a one-time policy exercise.
NIST’s August 2026 workshop summary reports participant discussion of governance challenges, attack surfaces, consistent AI taxonomy, the usability of risk guidance and AI-enabled defense. These are themes raised at a workshop, not measured estimates of how common a problem is. They reinforce the practical value of clear ownership, shared definitions and guidance that staff can apply. The NIST Cyber AI Profile workshop report summarizes those discussions.
Best Value
Which NIST guidance is active, and what is not a requirement?
Leaders can use current NIST material to inform risk management, but should distinguish a framework or project from a binding legal obligation. Applicable duties depend on jurisdiction and sector; the NIST pages cited here do not establish a complete list of legal requirements.
- NIST AI Risk Management Framework: NIST says AI RMF 1.0 is being revised. Its page also records an April 7, 2026 concept note for a trustworthy AI profile in critical infrastructure. The framework is guidance, not itself a law. See NIST’s AI RMF page.
- COSAiS: NIST describes this as an effort to develop implementation-focused overlays, using SP 800-53 and leveraging SP 800-218A, draft AI 800-1 and AI 100-2e2025. Its page includes proposed use cases and a predictive-AI discussion draft updated January 8, 2026. These are evolving project materials, not blanket binding requirements. See the COSAiS project page.
- Cyber AI Profile workshop: The August 2026 report summarizes a January 2026 workshop. Its themes reflect participant discussion and do not establish measured prevalence or a finalized requirement. See the workshop summary.
Use these sources to shape an organization-specific control program, and have qualified counsel assess applicable legal duties for the jurisdictions and sectors in which the organization operates.
A practical way to prioritize AI security reviews
When multiple AI systems compete for attention, compare them using consistent questions rather than a single generic “AI risk” label. The axes below synthesize NIST’s system and component framing with the operational themes in the SANS survey; they are not a formal NIST scoring rubric.
- System type: Is it an LLM assistant, predictive AI, a single or multi-agent system, or an AI development pipeline?
- Assets at risk: Which input, training, test or output data, model weights, configuration, software or hardware would matter if exposed, changed or unavailable?
- Access and autonomy: What systems and data can it reach, and can it take actions without human approval?
- Security objective and user impact: How could confidentiality, integrity, availability or user protection be affected?
- Operational readiness: Is ownership clear, is deployment maturity understood, have detection and response been validated, and are staff trained?
Prioritize reviews where sensitive assets, broad access or independent action coincide with weak ownership, untested detection or unclear response procedures. This approach directs effort toward the exposure and readiness of each system rather than assuming that every AI use carries equal risk.
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.
