Here’s a scene that’s made up but happening in plenty of offices right now. On Monday, a developer installs a coding agent on her laptop. By Wednesday, it can read three repositories, has found a cloud API key sitting in an environment file, and has permission to run shell commands. IT has no idea it’s there. Nothing about it looks like malware. It simply follows instructions, and anyone who manages to feed it the right ones gets to steer it.
Multiply that laptop by a few hundred staff and the old meaning of “endpoint” falls apart. It used to mean a laptop, a phone or a server. Today it also covers the agents, local models and automation scripts living on those machines, many holding far more access than whoever installed them realized.
That’s how endpoint detection and response (EDR), once one of the sleepier lines in a security budget, ended up in the middle of the AI conversation. If you’re deciding whether to run it yourself or hand it off, a guide to managed endpoint detection and response EDR from California IT provider Diamond IT lays out the basic deal. A managed service rolls 24/7 monitoring by a security operations team, proactive threat hunting, and incident response into a flat monthly fee. Running your own tooling means your own people watch the alerts. Either way, the five steps below still apply, written for teams juggling ordinary devices alongside a fast-growing crowd of AI processes.
Step 1: Get comfortable with a new kind of attack surface
The first change is in your head. Give an agent access to email, calendars, and the company wiki, and a cleverly worded prompt can talk it into leaking data it was never supposed to touch. Traditional antivirus often stays quiet, since no malicious file ever runs. The agent is just taking orders from the wrong person.
Security leaders know it. Darktrace’s State of AI Cybersecurity 2026 report, republished by the Cloud Security Alliance, found that 92% of security professionals worry about AI agents in their workforce and what they mean for security. Asked about their biggest concerns with AI in their own environments, they answered like this:
| Top worry about AI use | Share of security leaders |
| Sensitive data getting exposed | 61% |
| Data security and policy violations | 56% |
| Misuse or abuse of AI tools | 51% |
Rules are lagging, too. Only 37% said their company has a formal AI policy, a little lower than the year before. The report’s recommendation is refreshingly plain: treat every agent as an identity, give it the least access it needs, and watch what it does on the network and on endpoint devices. In practice, that means handling each agent the way you’d handle a server or a remote laptop, as a possible way in.
Step 2: Hunt down everything that’s running
Before anything gets bought or installed, take stock honestly. Laptops and servers are the easy bit. The harder list includes every API key, every agent framework, every scheduled job calling a language model, and every browser extension quietly shipping company data off to someone else’s AI.
Expect surprises. Agents tend to pop up on developer machines, inside CI pipelines, and in weekend side projects nobody signed off on. That’s shadow AI in action, and each one carries its own blast radius. Skip this mapping and the blind spots get built into your deployment before it even starts.
Step 3: Sort them by reach and freedom
Some agents barely matter. One that reads a public dataset is close to harmless. One that can write to a production database, push code, or email customers in someone’s name is another story entirely.
Give each endpoint two scores: how much it can reach, and how much it can do without a person approving each move. Whatever lands high on both gets the tightest monitoring, the fastest alerts, and the first sensors. That ranking ends up as the spine of your EDR policy.
A fair warning belongs here. EDR watches devices. It will spot the coding agent on a laptop, the local model on a workstation, and the script on a build server. An agent living entirely inside a SaaS platform never touches endpoint telemetry, though, so identity and SaaS logs have to cover that ground. Rank both kinds anyway, so nothing slips through just because it runs in the cloud.
Step 4: Roll out in waves, and lock down identities as you go
Choose a platform that understands non-human activity. Agents usually run under service accounts and move at machine speed. Without baselines tuned for that, the tool treats everything as suspicious and your analysts sink under the noise.
Then go in waves, starting with the riskiest endpoints from Step 3. Pushing sensors to a thousand machines at once and chasing false positives across all of them in week one is a reliable way to make your team resent the new tool.
Tighten identity at the same time. Rotate API keys on a schedule, keep permissions narrow, and log every agent action as though a privileged user took it, because these days one effectively did. The urgency comes from how cheap attacks have become. In the $50 AI hack, a single attacker sent a swarm of agents after more than 1,200 websites and got into around 100 companies, at roughly the price of a dinner out per attempt. Sensors on the machines where your own agents live are among the few controls placed to spot that kind of cheap, relentless probing early.
Smaller teams feel the value of a managed service most at this point. Someone in a security operations center is awake at 3 a.m. on a holiday weekend, when an agent abruptly starts pulling data it has never touched before.
Step 5: Tune for how agents behave, then keep at it
Once sensors are live, the work turns into vigilance. Build alerts around patterns that matter most for agents: a sudden surge in data moved, a call to an API it has never used, or a normally idle process springing to life at odd hours. A hijacked agent can act thousands of times in the time a person takes to click once, so how fast you catch it matters even more than with human users.
Response playbooks need a refresh as well. Isolating a laptop means unplugging it from the network. Isolating an agent might mean revoking a token, killing a process, or freezing a service account, and your team should already know which before the first real incident.
Finally, put tuning on the calendar. New agents arrive every quarter, frameworks shift, and permissions drift upward unless someone keeps trimming them. For most teams, a monthly review covers it: re-map endpoints, chase down orphaned agent accounts and stale keys, and reset baselines to match how things are used now.
Where this leaves you
EDR started as a way to catch malware on devices. Today it also has to keep watch over software that acts by itself, holding real permissions, inside your network. Teams that count agents as endpoints, map them honestly, and keep adjusting will stay ahead as AI tools spread. Everyone else tends to discover their agents during an incident.
So try one question at Monday’s stand-up. How many AI agents are running on company machines right now, and could anyone in the room list them all?
Related: The $50 AI Hack: How AI Agents Are Changing Cybercrime
