An AI security tool has just finished checking a suspicious file. It hasn't found anything malicious, so it decides the file is safe. Normally, that would mean the security team can stop worrying about it and move on to something else. And that's part of what makes automated malware detection so appealing. 

If AI can deal with harmless files on its own, analysts have more time to investigate real threats. But what if the AI missed something? Once a file has been cleared, it might never be checked by a person, even if it contains malware. That's where AI malware triage gets a little more complicated. 

em360tech image

It's one thing for AI to decide whether a file looks malicious. It's another for it to decide that nobody needs to investigate it any further.

What Does It Mean for AI to Clear a Suspicious File?

When a security tool checks a suspicious file, it's usually trying to find out whether the file contains malware. This is called malware classification, and the result is generally a label such as malicious or benign, which simply means harmless. Malware triage takes that a step further. 

Instead of just identifying what the file appears to be, the system also needs to decide what to do with it. There are three possible outcomes in an AI-assisted malware triage process:

  • Clear: The file appears harmless, so the system decides it doesn't need any more investigation.
  • Flag: The file appears malicious or suspicious enough for the security team to take action.
  • Escalate: The AI can't make a reliable decision, so it sends the file for further investigation.

That third option is especially useful because AI doesn't have to identify every possible threat to help security teams. It can still save analysts time by dealing with files it can assess reliably and passing the more difficult ones along. But there's a difference between an AI system saying it hasn't found anything malicious and actually having enough information to say a file is safe. 

For example, the AI might examine part of a program's code and find nothing suspicious. But if the malicious behaviour is hidden in a part it hasn't checked, the file could still be dangerous. So, before a business lets AI clear files without human review, it needs to understand how those decisions are made and what could happen when the AI gets them wrong.

Why Malware Detection Accuracy Doesn't Tell the Whole Story

It's easy to look at an AI malware detection tool's accuracy score and assume that a higher number means it's more reliable. If a model correctly identifies most of the files it checks, surely that's a good sign? Well, yes. 

But malware detection accuracy doesn't always tell you what kinds of mistakes the AI is making. And some mistakes can cause much bigger problems than others.

False negatives can leave threats uninvestigated

A false negative is when a security tool mistakenly identifies a malicious file as harmless. If a person is investigating the file, they might notice something suspicious and decide to take another look. But when AI is allowed to clear files automatically, there might not be anyone checking its decision. 

Depending on how the security system works, the file could be released from quarantine, allowed to run or simply removed from the list of files that need investigating. That's why security teams need to know how often their AI misses malicious files. This is called the false-negative rate, and it helps them understand how frequently dangerous files are being cleared by mistake. 

They also need to know which kinds of malware the AI struggles with. A model might be very good at recognising familiar threats but have a much harder time identifying malware that's designed to hide what it's doing.

Excessive escalation can erase the workload savings

There's another problem, though. What happens if the AI sends too many files for human investigation? If the system escalates every file it isn't completely sure about, it might avoid some dangerous mistakes. 

But analysts could end up checking almost as many files as they did before the AI was introduced. And plenty of those files might turn out to be harmless. So, security teams need to find a balance. They want AI to catch as many threats as possible without making analysts investigate hundreds of files that don't actually need their attention. 

A useful triage system should reduce the number of harmless files reaching analysts while still making sure suspicious ones get investigated. And researchers are starting to look more closely at how well AI handles these different decisions, rather than judging it on accuracy alone.

What TRIAGE-CPG Reveals About AI Malware Triage

A study published on 16 November 2026, called TRIAGE-CPG: A Cross-Model Study of Budgeted Static Malware Triage with Large Language Models, gives us a useful example of just how differently AI models can handle malware investigations. Researchers from Fraunhofer AISEC and Google developed a system that uses large language models to examine Windows programs. 

The system looks at decompiled code, which is code that's been converted from a finished program into a form that can be examined more easily. Rather than checking every part of a program, the AI starts with selected parts and can choose to examine more within a set limit. The researchers tested eight AI models using 1,000 Windows executable files, with 600 set aside for the final comparison. 

The best-performing model achieved 80.5% strict balanced accuracy and 85.5% triage balanced accuracy. These measure slightly different things. The first looks at how well the model classifies files, while the second also gives it credit for sending malicious files for further investigation instead of trying to make a final decision. 

That model correctly flagged 65% of malicious files and escalated another 12%. It also cleared 94% of the harmless files. But the interesting part is how differently the models behaved. Some sent relatively few files for further investigation, while others escalated or flagged far more. And reducing the number of files sent to analysts wasn't always safer. 

In one experiment, the researchers used a second AI model to review files that the first model had escalated. This reduced the number of files needing investigation from 44 to 35, but it also meant two more malicious files were missed. So, even though the second model made the investigation queue smaller, it also allowed more threats to slip through. 

There are some limits to what we can take from these findings. The researchers used equal numbers of malicious and harmless files, which isn't necessarily what a real security team would encounter. They also relied on agreement between VirusTotal detection engines to label files and excluded some cases where those engines disagreed. 

The tests focused on Windows executable files and examined their code without running them. So, the study gives us useful evidence about how AI handles malware triage, but it doesn't prove that any of these models can safely clear files in a real business environment.

When Is There Enough Evidence to Clear a File?

An AI model can be very confident that a file is harmless and still be wrong. Sometimes that's because it hasn't been able to examine everything it needs to make a reliable decision. So, security teams need to look at more than just the AI's final answer. They also need to understand what the system actually checked before reaching that conclusion.

What if the analysis is incomplete?

Malware doesn't always make it obvious what it's trying to do. Some malicious programs hide their code using techniques called packing or protection. Others can recognise when they're being examined by security tools and change how they behave. The MITRE ATT&CK guidance on sandbox evasion explains several ways malware can avoid being detected during testing. 

For example, it might check whether it's running in a virtual environment, wait for signs that a real person is using the computer or delay its malicious activity until after the test has finished. This creates a problem for AI malware detection. A file might look harmless because the system hasn't seen its malicious behaviour, rather than because the file is actually safe. 

The TRIAGE-CPG researchers ran into similar problems with packed and protected files. These were more likely to appear among the files that models got wrong or sent for further investigation. So, security teams need to know whether an AI has properly examined a file before trusting its decision. 

If the system couldn't read part of the code, doesn't support the file type or is missing important information, clearing that file could be risky.

Can AI recognise when it doesn't know enough?

One way to handle this is to have AI send files for human investigation whenever it isn't confident about its decision. That sounds sensible, but there's a catch. AI confidence scores aren't always a reliable indication of how likely the model is to be correct. A model can be completely confident about an answer and still get it wrong. 

Researchers call the relationship between a model's confidence and its actual accuracy confidence calibration. Put simply, they're checking whether the AI is as reliable as it thinks it is. A 2026 study published in *Array*, called Knowing When to Escalate: Prompt-Based Uncertainty Quantification for Selective Automation of LLM-Based Intrusion Detection, looked at this problem in AI systems that identify suspicious network activity. 

The researchers found that improving how models express uncertainty couldn't make up for poor classification accuracy. Their findings supported using uncertainty to help decide when to escalate cases, rather than relying on it to justify fully automated decisions. Although the study focused on network threats rather than malware files, the same concern applies. 

Security teams can't simply assume an AI knows when it's making a mistake. They need to test whether the system actually sends the cases it's most likely to get wrong for further investigation.

How Should Enterprises Evaluate Automatic Malware Clearance?

There's no single accuracy score that can tell every business when it's safe to let AI clear suspicious files automatically. A lot depends on the kinds of files the organisation deals with, how dangerous a missed threat could be and how much time its security team has to investigate suspicious activity. 

So, rather than relying on the model's overall accuracy, businesses need to test how well the entire malware triage process works in practice.

Test the decisions, not just the model

Are you enjoying the content so far?

The first step is to test the AI using files that are similar to the ones the organisation actually encounters. A benchmark with equal numbers of malicious and harmless files can be useful for comparing different models. But a real investigation queue might contain far more harmless files than malicious ones, which changes how useful those results are. 

Security teams should include known malware, harmless files, unfamiliar threats and files that are difficult to examine. They should also test what happens when information is missing or different security tools produce conflicting results. 

Then they need to look at how often the AI correctly clears harmless files, mistakenly clears malicious ones and sends uncertain cases for further investigation. It's also worth looking at the mistakes themselves. If the AI repeatedly misses malware hidden inside packed files, for example, the organisation might decide that those files should always be investigated by a person. 

And testing shouldn't stop once the system is introduced. Malware changes, AI models get updated and the files entering the system won't always stay the same. So, security teams need to keep checking whether the AI is still making reliable decisions.

Measure the work that reaches analysts

A smaller investigation queue might sound like a good result, but it doesn't necessarily mean the security team has less work to do. The AI might be sending fewer files to analysts, but those files could be much more complicated to investigate. Or it could still be sending through large numbers of harmless files that take time to check. 

To understand whether the system is actually helping, security teams should measure:

  • Escalation rate: How many files does the AI send for further investigation?
  • Queue quality: How many of those files turn out to be malicious?
  • Investigation time: How long do analysts spend checking them?
  • Missed threats: How many malicious files has the AI cleared by mistake?
  • Overall workload: Has the amount of investigation work decreased without creating unacceptable security risks?

These results need to be considered together. If the AI is saving analysts time by missing more malware, it isn't really solving the problem.

Define when automatic clearance is allowed

Once a business understands how reliably its AI makes decisions, it can start setting rules for when files can be cleared automatically. These rules should explain what information the AI needs before it's allowed to clear a file, along with the situations where further investigation is required. 

For example, a business might only allow automatic clearance for certain file types that the system can examine properly, provided the analysis is complete and the result meets a reliability level that's been tested. Files with missing information, conflicting findings or known analysis problems could be sent to analysts instead. 

The security team should also be able to look back at cleared files, see why the AI made its decision and investigate any mistakes discovered later. None of this guarantees that the AI will never miss malware. But it gives the organisation more control over which decisions are automated and helps it recognise when the system is no longer performing as expected. 

Ultimately, automatic malware clearance should happen because a business has tested and approved the conditions under which AI can make that decision, not simply because the AI says a file is harmless.

Final Thoughts: Safe Malware Clearance Depends on the Evidence

AI doesn't have to handle every malware investigation on its own to be useful. If it can reliably clear straightforward cases and recognise when something needs a closer look, it can help security teams spend more time dealing with genuine threats. 

But before allowing AI to clear suspicious files automatically, businesses need to know what the system has actually checked, where its limits are and how often it makes mistakes. And as malware and detection tools continue to change, those decisions will need to be tested again. The real question isn't just how much work AI can take away from analysts, but how much risk a business is willing to accept in return. 

As security teams work towards finding that balance, EM360Tech continues to share the research and practical insights that help businesses understand what emerging security technologies can actually deliver, and where they still have something to prove.