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 →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
Yes. Available evidence shows AI use is widespread while controls tailored to models, copilots and agents remain uncommon. In a 2024 UK survey, 68% of businesses used at least one AI technology, but only 40% of those users reported explicit AI-cybersecurity practices. Newer surveys show declining confidence in incident response, frequent agent permission failures and shortages of trained staff.
Adoption is broad, but AI-specific security is not
The UK Department for Science, Innovation and Technology (DSIT), working with IFF Research, surveyed 350 businesses in 2024. Sixty-eight percent said they were using at least one AI technology. General governance was common: 90% of all surveyed businesses had at least one governance or risk-management arrangement. That coverage did not translate into controls designed for AI behavior.
Among current AI users, 47% said they had no specific AI-cybersecurity practices and another 13% were unsure. Only 40% said they had explicit practices. The gap also varied by how AI was obtained: 31% of businesses buying off-the-shelf AI products reported AI-specific practices, compared with 63% of those outsourcing AI development.
| Measure | Finding | What it means |
|---|---|---|
| Businesses using at least one AI technology | 68% of 350 UK businesses (DSIT/IFF Research, 2024) | AI is already part of mainstream operations |
| All businesses with at least one general governance or risk arrangement | 90% (DSIT/IFF Research, 2024) | Most organizations have conventional governance structures |
| Current AI users with explicit AI-cybersecurity practices | 40% (DSIT/IFF Research, 2024) | AI-specific controls are the minority |
| Current AI users with no AI-specific practices | 47% (DSIT/IFF Research, 2024) | General security programs may not cover AI failure modes |
| Current AI users unsure whether they have such practices | 13% (DSIT/IFF Research, 2024) | Some organizations cannot identify their own AI controls |
General cybersecurity protects networks, identities and endpoints. AI security must additionally account for model and prompt changes, training and retrieval data, tool calls, autonomous actions, output reliability, model providers and the people authorized to approve or override an action. An organization can therefore pass a conventional security review while lacking a way to stop an agent from sending data, changing a record or making an unsafe recommendation.
#1 Best Overall
What counts as an AI incident?
The OECD definition reproduced in the Stanford AI Index 2026 treats an AI incident broadly:
“An event, circumstance, or series of events where the development, use, or malfunction of one or more AI systems directly or indirectly results in any of the following harms: (a) injury or harm to the health of individuals or groups; (b) disruption of the management or operation of critical infrastructure; (c) violations of human rights or breaches of legal obligations intended to protect fundamental, labor, or intellectual property rights; or (d) harm to property, communities, or the environment.”
That scope includes more than a hacked model. Examples include a prompt injection that causes unauthorized tool use, a data leak through retrieval or logging, a hallucinated decision that harms a customer, a deepfake used for fraud, model drift that changes outcomes, or an agent that exceeds its approved authority. A useful response plan must cover direct technical compromise and harmful consequences created by ordinary use or malfunction.
Recommended Free Tools
Incident-response confidence is deteriorating
Stanford’s AI Index 2026 reports McKinsey survey results for 2024 and 2025. The share of organizations reporting any AI incident stayed at 8% in both years, but affected organizations experienced incidents more frequently and rated their response capability lower.
Rank #2
| McKinsey measure | 2024 | 2025 | Change |
|---|---|---|---|
| Organizations reporting three to five AI incidents | 30% | 50% | Higher concentration of repeat incidents |
| Organizations reporting only one or two incidents | 42% | 29% | Fewer organizations remained in the lowest incident band |
| AI-incident response rated “excellent” | 28% | 18% | 10 percentage-point decline |
| AI-incident response rated “good” | 39% | 24% | 15 percentage-point decline |
| Response rated “needs improvement” | 13% | 21% | 8 percentage-point increase |
Risk perception rose at the same time. Concern about inaccuracy increased from 60% to 74%, and concern about cybersecurity from 66% to 72%. Organizations reported active mitigation for 71% of inaccuracy risk and 61% of cybersecurity risk. Those figures indicate that leaders recognize the problems, but recognition is not the same as a tested response capability.
Security teams are using AI cautiously and lack specialist capacity
A September 2025 SANS Institute survey found 81% of security professionals concerned about AI-powered threats. Operational use remained limited: 33% used AI to investigate incidents and 26% used it to respond. Sixty-six percent reported excessive false positives, a warning that adding AI to a security workflow can increase analyst workload unless outputs are validated and tuned.
Only 35% reported a formal AI risk-management and compliance program, while 65% said their teams needed more specialized AI and cybersecurity training. Adoption was concentrated in lower-complexity tasks such as alert enrichment and anomaly detection; investigation, code review and red teaming lagged. A practical implication is to treat AI-assisted security decisions as a supervised capability until accuracy, evidence retention and escalation behavior have been demonstrated.
Agents create authorization and visibility problems
AI agents can call tools, access business systems and take actions across multiple steps. The Cloud Security Alliance’s April 2026 report describes adoption that is decentralized and difficult to inventory:
Rank #3
- Forty-three percent of organizations said more than half of employees use agents regularly.
- Forty-four percent use two or three agent platforms, while 43% use four or more.
- Fifty-four percent reported between one and 100 unsanctioned agents.
- Only 15% said 76% to 100% of their agents had defined ownership.
Authorization failures were routine. Fifty-three percent said agents exceeded intended permissions occasionally or sometimes. Forty-seven percent had experienced an agent-related security incident, and 58% said detection and response took at least five hours. Only 13% felt highly prepared for upcoming AI regulations.
These findings make agent identity and runtime control central security requirements. Each agent needs a named business owner, a technical custodian, an approved purpose, an explicit tool and data allow-list, short-lived credentials where possible, and a human override or shutdown path. Static access granted during deployment is not enough when an agent’s plan can change at runtime.
Governance is shared, but accountability is often unclear
The International Association of Privacy Professionals (IAPP) reported in 2025 that 77% of organizations were working on AI governance, rising to nearly 90% among organizations already using AI. Responsibility was distributed rather than concentrated in security.
| Function identified as primarily responsible | Share of respondents (IAPP, 2025) |
|---|---|
| Privacy | 22% |
| Legal or compliance | 22% |
| IT | 17% |
| Data governance | 10% |
| Ethics or compliance | 6% |
| Security | 5% |
Half of AI-governance professionals surveyed worked in ethics, compliance, privacy or legal teams. Of 671 respondents, only 10 (1.5%) said they would not need additional staff in the next 12 months. The operating model implied by these results is federated: one accountable executive, named owners in security, privacy, legal, IT, data governance and the affected business unit, plus a defined escalation route when those interests conflict.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What an AI incident-response plan should include
1. Build an inventory before an incident
Maintain a current register of models, hosted services, copilots, agents, data connections, business purposes, versions, vendors, owners and user populations. Include systems acquired by departments without central approval. Record whether an application can read, write, execute code, send messages, make decisions or trigger physical processes.
2. Classify harm and severity
Use the OECD-style harm categories as the top-level test: injury or health harm, critical-infrastructure disruption, rights or legal violations, and harm to property, communities or the environment. Add operational severity levels based on affected people, data sensitivity, breadth, reversibility and regulatory notification duties.
3. Assign accountability and escalation
For every system, name a business owner who accepts the use case, a technical owner who can restrict or disable it, and contacts for security, privacy, legal or compliance, data governance and communications. Document who can authorize containment without waiting for a committee.
Free tools Windows power users keep installed
One-click scans. No signup required.
4. Enforce least privilege at runtime
Define the tools, records, data fields and actions an agent may use. Require approval for high-impact actions, separate read from write permissions, constrain destinations, expire credentials and re-check authorization on each consequential tool call. Keep a tested human-override and emergency shutdown procedure.
5. Preserve evidence
Where lawful and proportionate, retain prompts, model and policy versions, retrieved documents, tool calls, approvals, outputs, user identity, timestamps and downstream effects. Logging must allow responders to reconstruct what the system saw, decided and changed, not merely show that an API request occurred.
6. Define containment, recovery and notification
Specify how to disable an agent, revoke credentials, quarantine data, roll back a model or prompt change, correct affected records and notify customers, regulators or partners. Include a post-incident review that tracks root cause, control failure, recurrence risk and assigned corrective actions.
7. Exercise realistic scenarios
- Prompt injection that redirects an agent to an unapproved tool or destination.
- Data leakage through prompts, retrieval results, outputs or logs.
- Hallucinated or biased decisions affecting a person.
- Unsafe tool use, code execution or transaction approval.
- Model drift or a version change that alters behavior.
- Deepfake-enabled fraud or impersonation.
- Agent scope escalation caused by chained instructions or misconfigured permissions.
8. Train users and responders
Teach employees how to report suspicious outputs, verify high-impact recommendations, handle confidential data and recognize unsanctioned agents. Train responders on model behavior, prompt and retrieval evidence, agent traces and the legal duties that may apply to an AI-caused harm. The SANS finding that 65% of teams need more specialized training makes this a capacity requirement, not an optional awareness campaign.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errors9. Measure whether controls work
Track time to detect, time to contain, time to recover, repeat incidents, false-positive rates, percentage of systems with named owners, percentage of agents covered by permission policies, log completeness and the proportion of critical scenarios exercised successfully. Review these measures by business unit and system risk, not only as an enterprise average.
General cybersecurity versus AI-specific security
| Control area | General cybersecurity question | AI-specific question |
|---|---|---|
| Inventory | Which devices, identities and applications exist? | Which models, agents, prompts, data connections and versions exist, and who owns each? |
| Access | Who can sign in? | Which tools and data may an agent use for this particular action, and is authorization rechecked at runtime? |
| Monitoring | Was there an unusual login or network event? | What prompt, retrieval result, model version or tool call produced the behavior? |
| Response | Can the account or endpoint be isolated? | Can the model, agent, policy, credentials and downstream effects be stopped and safely reversed? |
| Governance | Which security team owns the control? | Who has authority across security, privacy, legal, data governance, IT and the affected business? |
Organizations do not need to replace their existing security program. They need to extend it with AI inventory, behavior and authorization controls, evidence that captures model-specific context, and rehearsed decisions for harms that may occur without a conventional breach.
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.
