Agentic Decompression Sickness
Jason Rebholz
Every previous computing revolution gave security teams 20–30 years to catch up. Agents gave us two.
At this point, hearing "things are moving so fast" triggers the same reflex that "unprecedented times" did during COVID. Yeah, we get it! But then I checked the dates…it was an “oh, shit…” moment.
Mainframe → PC (~30 years to “security maturity”): From the mid-1980s to the 1990s, PCs started to make their way into offices before Saved by the Bell went to reruns. Suddenly, the power of computing was in the hands of the general employee. Firewalls and anti-virus software became popular as PC demand grew. It wasn’t until 2013 that Anton Chuvakin of Gartner coined the term Endpoint Detection and Response (EDR), two years after CrowdStrike started.
On-premises servers → Hyperscalers (~20 years to “security maturity”): AWS launched in the mid-2000s as organizations sought to move away from managing their own data centers and use shared resources. It wasn’t until the mid-2010s that it became mainstream. From the security perspective, it wasn’t until 2020 when Wiz was founded and really accelerated cloud security as we know it today, much like CrowdStrike did for EDR! Let that sink in.
Human → agents (~2 years to adoption: security maturity pending): Cursor launched in 2023, and Anthropic released Claude Code in 2025. Engineers led the charge in using agents to offload tasks. Claude Code is widely regarded as one of the best tools available, and it was released a year ago. A year ago! With the release of Cowork in January 2026, the agentic capabilities Claude Code pioneered are now accessible to all employees. Security solutions are just emerging (Evoke Security 👋👋), while companies are just starting to think about what they need to do to secure this new layer.
Like a diver who surfaced too fast and got decompression sickness, every organization is feeling dizzy from everything moving so fast. The adoption is far outpacing any technology we’ve seen before. So, we look to the typical security playbook:
Policies, Risk Assessments, & Threat Modeling: For better or worse, most security starts with an acceptable use policy that no one can enforce. Then it shifts to a theoretical exercise to determine what risk the technology poses to the company and where things can go wrong.
Inventory: The first attempt to get our arms around the problem, figure out where the technology is in use. This leads to the first attempt to govern the technology.
Secure Configurations: A baseline of security, often popping up in how you configure the technology and identity access management. This then goes into drift monitoring for those controls. It’s the vulnerability management muscle applied to configuration management.
Run-time Monitoring and Blocking: This usually starts with tossing logs over a fence and maybe creating a few detection rules that are more for show than operational security. Then it matures to a true detection and response solution that can identify malicious or anomalous activity and block it before damage is done.
Here’s the problem, though. With “how fast things are moving,” you can’t afford to wait 15+ years for security solutions to mature. That includes not waiting for the incumbent security providers to figure out how to tackle agent security beyond their latest marketing attempts. If you’re up for a good scavenger hunt, go look at all the leading security companies that claim they do AI detection and response. Spoiler alert: they’re still talking about LLM firewalls and looking at prompts, not actions. That is so 2025. Remember how fast things are moving?
Here’s what you need in a solution today:
Inventory and Governance: See every agent. But it’s not enough to know that an agent exists. You need visibility into which tools it can call and which data it can access.
Identify Risky Configurations: Surface over-permissioned agents or poor configurations that can introduce security gaps. If your agent isn’t using a permission, get rid of it. Force it back to secure defaults.
Enforce Secure Policies: Establish secure policies at the organization, department, employee, endpoint, and agent levels. Your engineers need different permissions and controls than your finance team. This extends from the agents they can run, all the way down to the actions an agent can take in a specific tool.
Run-time Detection & Response: Get full visibility into what every agent is doing. Every tool call, every action. Monitor and block malicious or anomalous activity. Don’t rely on human-in-the-loop to keep you safe.
Most agentic security solutions solve one of these. Evoke solves all of them. Skip the agentic decompression sickness and go straight to securing your agents with Evoke. Schedule a demo here.