Lesson 5 of 8
Agents, agentless and production safety
Whether a tool needs agents decides what it can test, how long deployment takes and how much your operations teams will worry. Safety controls decide whether the pilot survives its first week.
What an agent is for
An agent is software you install on a host so the validation tool can act from inside it. Vendors call them different things: Cymulate has test points, SafeBreach has simulators, Validato has Validators, Mandiant Security Validation has Actors, SCYTHE uses agents or implants. The job is the same.
From inside a host, a tool can do what malware does after it lands: run processes, touch memory, write files, make outbound connections. That is the only way to test an endpoint control directly. If you want to know whether your EDR stops a technique, something has to attempt the technique on an endpoint the EDR protects.
Agents also let a tool test network controls from a realistic position. An agent in the finance VLAN can test whether the firewall lets finance hosts reach the internet over unusual ports.
What agentless testing can do
Agentless tools act over the network, through APIs, or from a single appliance or container you provide. Automated pentesting tools are mostly agentless. Horizon3.ai describes NodeZero internal tests running from a customer-provided Docker host or OVA, and external tests from its own cloud. A Pentera guide describes its Core, Surface and Cloud suite as agentless.
From the network, a tool can discover hosts, find services, try exploits and credentials, and move between systems once it has access. That is how real attackers spend most of an intrusion, so agentless testing is a good fit for path questions.
Some control tests also work without endpoint agents. Email gateway tests send messages to a mailbox. WAF tests send requests to a web application; SafeBreach notes its WAF validation is agentless. AttackIQ describes its Flex tier as agentless: it runs self-contained test packages, on Windows only, with a smaller content set than its Enterprise tier.
The trade-off
Agents cost time. Each one needs a change request, a deployment, and an exception in the EDR so the agent is not quarantined. Some teams cannot install them at all on certain systems: medical devices, OT, servers run by a third party.
Agentless costs coverage. Without code on the endpoint, a tool cannot grade an endpoint control technique by technique. It sees only whether its path was blocked.
Many vendors now offer both modes. Picus documents agentless and agent-based attack path assessments. Cymulate says one test point per environment is enough for its internal tests. The finder's third question, whether you can install agents, splits the types along this line.
What production-safe means
Every vendor on this site says its tests are safe to run in production. The claim covers different things.
For breach and attack simulation, safe usually means the scenario imitates malicious behaviour without the harmful part: a binary that behaves like ransomware but only encrypts test files it created, or an exfiltration test that sends fake records. The control sees realistic behaviour; nothing real is damaged.
For automated pentesting, safe usually means real exploits with guardrails. Pentera lists throttling, impact limits, an emergency stop, read-only modes and audit logs among its controls. The tool really does use the vulnerability; the guardrails limit what it does next.
Those are different levels of risk, and both can be acceptable. What is not acceptable is not knowing which one you are running.
Questions to settle before the first run
- What does each class of test change on a host: files, registry, services, accounts?
- How does the tool clean up, and how do you verify the clean-up?
- Can it lock out accounts, for example through password spraying? How is that limited?
- Does it ever exploit a vulnerability for real, and can you turn that off per target?
- How fast does it scan, and can you throttle it?
- How do you stop a run, and how long does the stop take?
- What is logged, where, and for how long?
- What happens if the agent or appliance loses its connection mid-test?
Ask for these answers in writing, from the vendor's documentation where possible. A vendor that cannot answer clearly is telling you something about how its product is run.
EDR exceptions
A detail that trips many pilots: to install an agent, you may need to tell the EDR to ignore it. If the exception is too broad, the EDR ignores the agent's test activity too, and every endpoint test passes. Check with the vendor exactly which exclusions it needs, and make them as narrow as the documentation allows.
What to take from this lesson
Decide where you stand on agents before you talk to vendors. If you cannot install them, control testing on endpoints is mostly off the table and automated pentesting moves up. If you can install them on some hosts, choose those hosts as your starting points. Then make the safety questions a condition of the pilot, not a footnote to it.
Previous: Scoping your first validation. Next: Turning findings into fixes.