To combat this, security teams must move beyond simply telling employees what not to do. Practitioners should prioritize the implementation of AI firewalls or reverse proxies that can intercept prompts in real-time, using Data Loss Prevention (DLP) tools to scrub PII or proprietary code before it reaches a third-party LLM. Identifying shadow AI also requires a rigorous discovery audit of web proxy and firewall logs to see which domains are being accessed. By providing a sanctioned, enterprise-grade AI catalog (a curated list of approved tools with preconfigured guardrails), security can offer a safe path for employees to meet their workload demands without compromising the perimeter.

“The mandate for 2026 is clear. The gap between adoption and readiness is a direct threat to digital resilience.”

This lack of control extends into the most critical area of security operations: incident response. The data shows a staggering lack of operational readiness regarding a kill switch — the ability to halt a system during a crisis. Over half of organizations — 56% — cannot say how long it would take to halt an AI system during a security incident, and a mere 12% have actually tested a process for a controlled shutdown. Practitioners must define specific technical protocols for revoking API keys and isolating autonomous agents. These steps must then be validated through dedicated AI-risk tabletop exercises that simulate specific threats like deepfakes or prompt-injection attacks.

The disconnect between risk recognition and mitigation is equally concerning. Practitioners cite privacy violations and social engineering as top concerns, yet only 45% of organizations treat AI risk as an immediate priority. This is compounded by an erosion of ethical focus. Only 11% of practitioners today believe their organizations are giving sufficient attention to ethical standards in AI implementation. While this represents a sharp drop from the 41% reported in 2025, it is important to note that the survey framing differed between years; regardless of the methodology shift, the directional signal is clear: ethics are being sidelined in the rush to deploy. We must bridge this by mandating human-in-the-loop requirements for high-stakes AI decisions and updating Privacy Impact Assessments (PIAs) to ensure we know exactly where user prompts are stored and whether they are being used to retrain a vendor’s global models.

Key Actions for Security Teams

1. Operationalize the Kill Switch

Security teams must document specific technical protocols for revoking API keys and isolating autonomous agents during a breach. These procedures must be validated through targeted AI-risk tabletop exercises to ensure the organization can move from theoretical policy to a functional, tested shutdown capability.

2. Build Guardrails, Not Just Policies

Move from blocking to enabling by auditing web proxy and firewall logs to identify unvetted AI usage and replacing those tools with a sanctioned enterprise-grade AI catalog. Implement technical enforcement, such as AI firewalls and DLP-enabled proxies, to automatically scrub PII and intellectual property from prompts.

3. Reframe Governance and Accountability

Broaden governance beyond the security team to address the board-level confidence gap by reporting on risk avoidance metrics rather than productivity ROI. These must include technical requirements for human-in-the-loop reviews of high-stakes AI decisions. Updated vendor procurement standards should explicitly address data-retraining policies — ensuring corporate prompts are not used to train third-party global models.

Security leaders must address the tone at the top. With only 38% of practitioners confident that their boards are acting on AI risks, and only 16% of organizations seeing their ROI expectations met, the narrative around AI must change. Rather than trying to prove unproven financial gains, security practitioners should report on risk avoidance metrics — highlighting the volume of sensitive data intercepted by guardrails or the number of unauthorized tools blocked.

The mandate for 2026 is clear. The gap between adoption and readiness is a direct threat to digital resilience. Closing it requires us to move beyond writing policies and start building the technical infrastructure that ensures if an AI system must be stopped, we have the power to stop it. Our role is to ensure that as the organization accelerates AI deployment, it does not leave its security, its ethics, or its operational control behind.