Share with your CISO
Nearly half of all AI agents running in enterprise production today operate without meaningful security controls, according to new research on agentic AI security gaps. The NeuralTrust State of Agentic AI Security 2026 report puts the figure at 48% of actual agents, not companies lacking a policy. Between December 2025 and April 2026, enterprise agent fleets doubled while monitoring coverage held flat at 52%. IBM pegs the average cost of an AI agent-related breach at $4.7 million, and 88% of enterprises with deployed agents have already experienced at least one security incident.
What this means for your business
The organizations most exposed here aren’t the ones that skipped a policy meeting. They’re the ones that moved fast on agents, which is most of them. Gartner finds 38% of large organizations now run fleets of more than 100 agents, and Deloitte finds nearly 80% lack mature governance for those fleets. If your security team can’t enumerate every agent in production, you’re already in the category the research describes. The question isn’t whether you have a gap; it’s how many agents are in it.
The structural problem the report identifies is real and worth sitting with. Application security assumes the code you reviewed is the code that runs. Agents break that assumption completely. A prompt injection attack, where adversarial instructions are embedded in a document or email that the agent retrieves mid-task, doesn’t touch your code at all. It arrives at runtime, after every pre-deployment control has already passed. Thirty-four percent of deployed enterprise agents have already been hit by this. The attack surface lives in what the agent can reach, not in what you shipped.
The shadow agent problem compounds everything. The recurring failure mode looks like this: a prototype built over a weekend to solve a real problem quietly becomes a production system because nobody ever made a formal decision to promote it. The security review never happened because it was “just a test.” Months later it’s handling live customer data with permissions scoped for exploration, not production. You can’t audit what you don’t know exists, and most security organizations have no reliable inventory of their agent fleet, let alone behavioral telemetry, the real-time data stream showing what each agent is actually doing at runtime.
NeuralTrust is a vendor with a product to sell into exactly this gap, which tilts their framing toward urgency and makes the 48% figure worth treating as directionally accurate rather than precise. But the underlying dynamic doesn’t require their numbers to be exact to be actionable. If your next budget conversation doesn’t include a line item for agent discovery and runtime monitoring, you’re defending a perimeter that no longer matches the attack surface your organization actually has.
Concept deep-dive: Prompt Injection
Prompt injection is the technique of embedding malicious instructions inside content that an AI agent reads and acts on, causing it to deviate from its intended behavior. Think of it as a forged memo slipped into a stack of legitimate documents: the agent can’t tell the difference between instructions from its operator and instructions hidden in a file it retrieved. Indirect injection, where the payload arrives through a third-party source rather than the user, is nearly impossible to catch with static code review.
Based on reporting from 48% of Your AI Agents Are Running Without Meaningful Security Controls, originally published 2026-09-12 05:12:00.
