AI is Exposing EDR Defenses. Here's How to Close the Gap.

An autonomous AI agent recently reverse-engineered the internal detection logic of a leading commercial EDR platform.

Not a nation-state lab. Not a team of veteran reverse engineers working for months. A single agent with read access to the software and a reverse-engineering tool did it in hours.

The research comes from SpecterOps and what they found isn't specific to one vendor; it's a warning for anyone relying on EDR.

AI Can Expose How EDR Detects Threats

SpecterOps set out to test whether an AI agent could reverse-engineer a leading commercial EDR platform's on-host detection logic.

It did, in hours, instead of the months that kind of analysis would normally take a human team.

The project, which SpecterOps calls "Day Shift," paired an AI model with Binary Ninja, a reverse-engineering tool, connected through the Model Context Protocol (MCP). The agent worked against the platform with minimal human guidance, and moved through two model generations over the course of the research.

It found the platform's entire detection arsenal sitting in plain sight.

SpecterOps verified each of these findings live against the running platform.

How Attackers Evade EDR with Stolen Detection Rules

Once the rules are extracted, an attacker doesn't have to guess what will trigger an alert.

Any detection logic that runs in a space the adversary controls can be enumerated and evaded is problematic. Because those rules already exist as files on the compromised machine, an attacker can study them offline and plan around them in advance, without doing anything that would tip off the EDR platform.

Here's what makes this worse: it doesn't take admin-level access. Read-only access to the machine was enough to pull the full set of rules, the machine-learning thresholds, and the list of paths and actions the platform treats as safe.

SpecterOps found a real example of what makes this possible.

Saving a copy of the SAM file – normally a strong sign of credential theft – went undetected because it used a file path the platform had been told to ignore. Once an attacker can see that allowed list, they don't need to guess which paths are safe. They already know.

EDR isn't Broken, it's Just One Layer

EDR remains essential for endpoint prevention, response, and forensics. This isn't a case against EDR as a category; it's a specific finding about one layer within it.

On-host static and behavioral rule sets are now provably extractable using commodity AI tooling without alerting anyone, so that layer alone can't be the thing carrying your detection strategy.

Any detection strategy that leans entirely on the endpoint has a blind spot the moment that endpoint is compromised.

The fix isn't abandoning EDR. It's making sure you have a layer that keeps working when the EDR has been evaded.

What used to gate this kind of analysis was expertise. The small pool of reverse engineers capable of doing this work by hand. This latest research shows that gate is now a script and an AI model with access to a disassembler.

The Network is the Layer That Doesn't Fail

Network telemetry isn't a file sitting on the compromised host, so there's nothing for an attacker to extract, decrypt, or reverse-engineer evasions against ahead of time.

That's the structural difference. There's no ruleset to pull down and study at leisure because the detection logic isn't stored anywhere the attacker can reach.

Instead of relying on the endpoint, modern NDR captures what actually happened on the wire. Real SMB access to sensitive files, real command-and-control beacon behavior, real lateral movement between hosts.

In the SAM hive example, a registry save through an allowlisted path still generates network behavior. Network behavioral models see that traffic regardless of whether the local EDR ruleset decides to allow it.

Close the Detection Gap AI Just Exposed

For us at ExtraHop, this isn't a call to lose all your faith in EDR. It's a call to stop treating it as sufficient on its own.

On-host static and behavioral rules can now be extracted by commodity AI tooling in hours, so that layer alone can't carry the full weight of your detection strategy. A network-derived layer that doesn't rely on just on endpoints is the piece that closes the gap.

Learn how threat actors are evading EDR. Read EDR Evasion 101: 29 Ways Attackers are Slipping Past Defenses.