88% of organisations have had confirmed or suspected AI agent security incidents in the past year. A new set of industry reports shows that most enterprises are still deploying autonomous agents faster than they can secure them.

How fast agents are being adopted
AI agents are no longer simple chatbots. They run multi-step workflows, query databases, call APIs and can even create other agents. Gartner expects 40% of enterprise applications to embed AI agents by the end of 2026, up from 5% in 2025, an eight-fold increase in one year.
Deployment is already well under way. One recent survey found that 80.9% of teams have moved past planning into testing or production, and more than half of companies use retrieval-augmented generation or agentic pipelines. Agents now handle customer data, carry out business transactions and make decisions that affect real operations.
Yet only 14.4% of organisations say all their AI agents went live with full security and IT approval.
What organisations cannot see
Only 21% of executives have complete visibility into agent permissions, tool usage or data access. The average enterprise is estimated to have 1,200 unofficial AI applications in use, and 86% of organisations report no visibility into their AI data flows.
Staff behaviour adds to the problem. Research shows 63% of employees have pasted sensitive company data, including source code and customer records, into personal chatbot accounts. Unofficial apps combined with unmonitored data flows leave blind spots that traditional security tools were not built to find.
The cost is measurable. Shadow AI breaches cost an average of $670,000 more than standard security incidents. Among companies with more than $1 billion in annual revenue, 64% have lost more than $1 million to AI failures.
Confidence that the controls do not support
Researchers describe a "confidence paradox". 82% of executives are confident their existing policies protect them from unauthorised agent actions. The technical findings point the other way:
- 80% of surveyed organisations reported risky agent behaviour, including unauthorised system access and improper data exposure
- 45.6% still rely on shared API keys for agent authentication
- only 21.9% treat agents as independent identities with their own credentials
- 27.2% use custom, hardcoded logic for authorisation decisions
- 25.5% of agents can create and task other agents without further approval
That gap between perceived and actual protection gives a false sense of readiness while real vulnerabilities accumulate.
The attacks are real. Prompt injection has moved from research papers into production incidents, because language models cannot reliably tell legitimate instructions apart from malicious content they are asked to process. In one documented case, a supply chain attack on a plugin ecosystem harvested agent credentials from 47 enterprise deployments. The attackers used them to reach customer data, financial records and proprietary code for six months before anyone noticed. Compromised agents can run unauthorised commands, exfiltrate data and move laterally across systems. In healthcare, with its sensitive patient data, reported incident rates reach 92.7%.
Where to start
Slowing adoption is unrealistic at this point. The workable approach is to build security into how agents are developed and deployed.
Give each agent its own identity. Agents should have their own credentials with scoped permissions, like human users or service accounts. Shared API keys and hardcoded authorisation logic add risk for no benefit.
Set controls by risk. An agent that reads public documentation needs lighter controls than one that can change production databases. Match the controls to what the agent can actually do.
Make guardrails the default. Sandboxed execution, scoped credentials and audit logging should come built into the platform, without anyone having to request them.
Red-team continuously. Automate attack testing in the deployment pipeline instead of waiting for production incidents to reveal vulnerabilities.
Enforce policy at runtime. Decide what each agent may do before it goes live, and have the system enforce those limits on every action.
For organisations in the Gulf that are earlier in their agent rollouts, this is an advantage: building secure agent infrastructure from the start is easier than retrofitting it. A useful first step is to ask whoever owns your AI programme one question. How many agents are running in our organisation today, and which credentials does each one use? If nobody can answer with a list, visibility is the first gap to close.