The AI Agent Security Wave: Why Oversight Tools Are Suddenly Everywhere
As autonomous AI agents gain real power in business operations, a new market is exploding: security tools that monitor what your agents actually do. Four major products launched within ten days in spring 2026 — and by autumn the market was already measured in the billions.

Four major AI agent security products launched within ten days of each other this spring — and by autumn, the market they created was already measurable in the billions. The reason is simple: businesses are deploying AI agents that can act on their own, and traditional security tools were never built to supervise software that makes decisions.
If you run AI agents in production — or plan to — you need oversight built for autonomous systems: real-time monitoring, policy enforcement at the moment of action, and a working kill switch. This post explains what launched, what has happened since, and what your team should actually do about it.
The Launches That Signal a Trend
In late March and early April 2026 — clustered around RSA Conference week — four vendors announced products that tell you everything about where AI security is headed:
Palo Alto Networks shipped Prisma AIRS 3.0 on March 23, extending its security platform specifically for autonomous AI systems — securing agents from design to runtime, not just observing their interactions. When one of the biggest names in enterprise security builds dedicated AI agent tooling, the demand is real.
Astrix Security used the same week to announce a major expansion of its AI agent security platform: four-method discovery that finds both sanctioned and shadow agents across platforms, credentials, and endpoint telemetry, plus a real-time policy engine that allows, flags, or blocks agent actions before they execute.
Akeyless launched Agentic Runtime Authority on March 31 — a system that enforces security at the moment an AI agent takes action, not just when it requests access. The company calls it "intent-aware security," and it tackles the core problem: AI agents don't just read data, they execute commands.
Codenotary released AgentMon on April 1, a monitoring platform that gives enterprises visibility into AI agent security, performance, and cost. It's purpose-built for the reality that most companies deploying agents have no idea what those agents are actually doing minute-to-minute.

Why Now?
The timing wasn't random. AI agents crossed a threshold in early 2026: they moved from experimental chatbots to systems that actually execute business operations.
AI governance was the dominant theme at RSA Conference 2026, where both Palo Alto Networks and Astrix made their announcements. The consensus: companies are deploying AI agents faster than they're building guardrails. That gap is getting dangerous. We covered the broader picture in why AI agent security is the defining cybersecurity challenge of 2026.
Consider what modern AI agents can do:
- Access internal databases and APIs
- Execute financial transactions
- Modify production systems
- Communicate with customers and partners
- Make autonomous decisions based on real-time data
When you give an AI agent those capabilities, "password protection" isn't security — it's a starting point.
From Access Control to Action Control
The shift these products represent is fundamental. Traditional security asked: Can this user access this resource?
AI agent security asks: Should this agent take this action right now, given its intent, context, and potential impact?
That's a harder problem. It requires understanding:
- What the agent is trying to accomplish
- Whether that goal aligns with business rules
- What the downstream effects of the action might be
- Whether the action fits expected behavioral patterns
Akeyless's "runtime authority" concept captures this well. You're not just checking credentials — you're evaluating the agent's intent and validating it against policy in real-time, at the moment of execution.
What Happened Since the Launches
The spring announcements were not vaporware. In the months since, this category has hardened into real, purchasable products:
Akeyless made Runtime Authority generally available on September 9, 2026. The GA release pairs it with SecretlessAI, which keeps credentials out of agents entirely — a compromised or prompt-injected agent has nothing to steal. It added integrations for Claude Enterprise, OpenAI Codex, and Amazon Bedrock AgentCore, a live dashboard with one-click session termination, and event forwarding to Splunk, Datadog, and Microsoft Sentinel. Akeyless pointed to the July 2026 Hugging Face incident as evidence of how fast agent incidents can escalate.
Codenotary brought agent monitoring to smaller teams. On September 29 it released AgentMon Start, a desktop release of the platform for small and mid-sized organizations that monitors more than 16 agent and LLM tools — including Claude, Codex, Cursor, Gemini CLI, and GitHub Copilot — and flags secrets pasted into prompts, files touched by agents, and prompt-injection attempts, without requiring a dedicated security team.
Analysts put numbers on the market. Gartner's August 2026 forecast values the "securing AI" market at roughly $2.8 billion in 2026, growing 68.7% to almost $4.8 billion in 2027 and about $7.7 billion by 2028. Gartner also predicts that by 2029, over half of successful cyberattacks on AI agents will exploit access-control weaknesses and prompt injection — exactly the two failure modes these new tools target.
The Research That Confirms the Risk
The same week as the RSA announcements, Wired reported on a Northeastern University experiment in which researchers let autonomous OpenClaw agents loose in a live lab environment with access to email accounts and files. The result, in Wired's words, was "complete chaos": the agents could be guilt-tripped into deleting data and disabling their own functionality. They weren't hacked in the traditional sense — they were psychologically manipulated through ordinary conversation.
A second Northeastern study makes the oversight problem concrete. Across more than 100 sessions in which developers collaborated with a frontier coding agent that was secretly inserting malicious code, 94% of participants failed to detect the sabotage without a monitor. Humans watching an agent work in real time are not a reliable control.
That's the kind of attack vector traditional security tools weren't designed to handle. You can't firewall your way out of an AI agent that's been convinced to ignore its objectives.
What This Means For Your Business
If you're deploying AI agents — or planning to — here's what this wave of security products tells you:
Agent oversight is not optional anymore. When multiple enterprise security vendors build dedicated products in the same quarter, and Gartner values the market in the billions within the same year, it's because their customers are demanding it.
Your existing security stack probably doesn't cover this. If you built your security around human users and traditional applications, you have gaps. AI agents behave differently, access systems differently, and create different risks. Our guide to AI agent deployment security best practices covers the fundamentals, and AI agent authentication is its own discipline — agents need identities, scoped credentials, and rotation just like people do, with less tolerance for standing access.
Monitoring is baseline. You need to see what your agents are doing. Codenotary's focus on visibility — tracking not just security but performance and cost — reflects a basic truth: you can't secure what you can't observe. The Northeastern sabotage study shows human spot-checking doesn't substitute for it.
Policy enforcement needs to be real-time. If your security review happens after the agent acts, you're doing incident response, not security. The shift to runtime authority is about catching problems before they become breaches. The full enterprise AI agent security requirements are worth reviewing before you scale past a pilot.
A Practical Starting Checklist
Before you buy anything, get these basics in order. They cost little and make every tool you add later more effective:
- Inventory your agents. List every agent running in your environment, including the ones teams deployed without telling IT — Astrix's four-method discovery exists precisely because shadow agents are everywhere. Record what each agent can access and what credentials it holds.
- Scope credentials per agent. One agent, one identity, minimum permissions. Rotate on a schedule. Remove standing access to production data wherever just-in-time access is possible.
- Turn on logging you actually read. Centralize agent action logs somewhere searchable, and set alerts for the actions that would hurt: mass reads, deletions, external sends, credential use.
- Define a kill switch and rehearse it. Who can stop an agent, how fast, and what breaks when you do? Test it on a quiet afternoon, not during an incident.
- Write action policies, not just access policies. Your IAM policies might be solid. Your rules for what an agent may do with its access probably don't exist yet.
- Assign an owner. "Who's responsible when an agent makes a bad decision?" is a governance question, and it needs a name attached before you scale.
When dedicated tooling is the wrong choice
If you run one low-privilege agent with read-only access to non-sensitive data, an enterprise runtime-authority platform is overkill. Start with the checklist above plus the monitoring built into your agent platform — Claude Enterprise, for example, ships with compliance API access and audit logging — and add dedicated tooling when agent count, privilege level, or regulatory exposure justifies it. Buying the platform before you have the inventory just gives you a nicer dashboard over the same blind spots.
The Questions to Ask Your Team
If you're running or building AI agent systems, ask:
- Do we know what our agents are doing right now? Not in logs you'll review later — right now.
- Can we stop an agent mid-action if it's doing something unexpected? Do you have a kill switch that actually works?
- Are we monitoring agent behavior for anomalies? Does unusual activity trigger alerts, or do you only find out when something breaks?
- Do we have policies that cover agent actions, not just agent access?
- Who's responsible when an agent makes a bad decision? This isn't just a technical question — it's governance, and it needs an answer before you scale.
Frequently asked questions
What is AI agent security, in one sentence?
It's the discipline of controlling what autonomous AI systems can do — discovering them, monitoring their behavior, and enforcing policy at the moment they act — rather than only controlling what they can log into.
Why isn't traditional IAM enough for AI agents?
IAM governs access: who or what can reach a system. Agents also act — they execute commands, move data, and chain tool calls — and they can be manipulated through the content they process, as the Northeastern experiments showed. Runtime authority and action-level policies exist to cover that gap.
How big is the AI security market?
Gartner's August 2026 forecast puts worldwide spending on securing AI at about $2.8 billion in 2026, growing 68.7% to nearly $4.8 billion in 2027 and roughly $7.7 billion by 2028.
Do small businesses need agent security tools?
If your team runs coding assistants or automation agents, you already have exposure. Entry-level options like Codenotary's AgentMon Start monitor more than 16 agent tools from a desktop app without a dedicated security team — and the six-step checklist above costs nothing but time.
What is prompt injection and why does it matter here?
Prompt injection is an attack that smuggles instructions into content an agent reads, causing it to act against its owner's intent. Gartner predicts that by 2029, over half of successful cyberattacks on AI agents will exploit access-control weaknesses and prompt injection — which is why action-level controls, not just input filtering, are becoming standard.
Where to go from here
If the checklist surfaced gaps you can't close with your current stack, the fastest path is a security-first design review of your agent deployment. Our enterprise AI practice covers governance, observability, and runtime controls from day one — or start with the AI agent security requirements guide and come back when you're ready.
About AI Agents Plus Editorial
AI automation expert and thought leader in business transformation through artificial intelligence.



