AI predictions for DFIR professionals

It is undeniable that AI is changing the way security professionals work, and the DFIR field is no exception. Tasks that used to take hours, such as writing a parser for an obscure artifact or summarizing thousands of event log entries, can be done now in minutes. With that said, I decided to write this post with some personal predictions for those of us working with DFIR.

Prediction 1: AI will fully automate routine DFIR tasks

A typical investigation follows well-known steps: 1) collect a triage package; 2) parse the artifacts; 3) build a timeline; 4) hunt for anomalies; and 5) write the report. Most of the first steps are mechanical. We run the same collectors and the same parsers, and we pivot on the same fields, case after case.

I expect this whole pipeline to run on its own. An agent will receive the collections, parse every artifact, correlate activity across hosts, and deliver a first draft of the report (containing initial access vector, first compromised machine, lateral movement indicators, persistence, and so on) before an analyst even opens the case. In fact, I have had this idea since I started working in DFIR, that's why I developed Triager. The first version was only CLI, I later adjusted it to a GUI with AI-enabled features. Right now, it can automate much of my work, but still needs to be supervised by a human.

The catch is traceability, since a hallucinated sentence in a forensic report can send an investigation in the wrong direction or not be accepted in court. The tools that win will be the ones where every conclusion points back to the correct artifact, record, and timestamp it came from, so a human can verify it in seconds. I strongly believe that we are living in a period of transition: we are supervising results generated by AI, this may last 2, 5, or even 10 years. However, inevitably, we will reach a point where AI will achieve reliable results that will not even require human verification.

Prediction 2: AI will automatically analyze malware

LLMs are already surprisingly good at reading assembly (not to say almost perfect analysis of decompiled code), renaming functions, and explaining what a routine does. With MCP servers for IDA and Ghidra, a model can navigate a binary on its own and perform operations such as follow cross-references, decrypt strings, and extract configurations.

My prediction is that the first-pass report for commodity malware (capabilities, C2 configuration, IOCs, and a draft YARA rule) will be fully automated. Loaders, stealers, and RATs that are variations of known families will rarely need a human reverse engineer. Projects like REA: Reverse Engineer Anything can help us deconstruct both benign and malicious software to better understand its behavior.

Attackers will adapt too. In 2025, Check Point reported a sample with an embedded prompt injection trying to convince AI analysis tools that the file was benign. It did not work in the researchers' setup, but it will not be the last attempt. Just as we have sandbox evasion, we will have AI evasion, and analysts will need to recognize when the model is being fooled.

Prediction 3: Many tools will arise, with different purposes

The cost and time of building a tool has dropped dramatically. A parser for a new artifact, a GUI feature for a command-line tool, a script that glues three APIs together: all of this can now be done in an afternoon (or even less). We will see an explosion of DFIR tools, many of them built for a single case and abandoned right after.

This is both good and bad. Good, because niche artifacts will finally get parsers. Bad, because a tool that silently misreads a timestamp or skips a record is worse than no tool at all. Also, depending on the tool you are developing, AI may generate a lot of code that will never be fully read, tested or understood. I expect the community to start valuing correct generalization more than features: test datasets, baselines, and comparisons against established tools from the pre-AI era.

Prediction 4: Attackers will move faster, and so must we

Everything I said about defenders also applies to attackers. AI lowers the effort for writing custom tools, crafting convincing phishing in any language, and automating reconnaissance, exploitation, and lateral movement. The time between initial access and impact is already short, and I expect it to shrink even more.

This has two consequences for DFIR: first, malware generated on demand for each victim will make signature-based detection harder, pushing us further toward behavior and TTP-based detection. Second, response time will matter more than ever. If the attacker operates at machine speed, a response that waits for an analyst to manually parse the evidence will always be late (AI wars).

Prediction 5: AI will become evidence

Companies are deploying AI agents with access to e-mail, file shares, code repositories, and internal APIs (not to mention our chats). Sooner or later, one of them will be the entry point or the exfiltration channel of an incident, whether through prompt injection, a malicious MCP server, or simply stolen API keys.

When that happens, we will need to investigate a new set of artifacts; for example: prompt histories, agent configuration files, and MCP server definitions. In criminal cases, the usage of AI leaves interesting traces in RAM that need to be analyzed. Most organizations today do not log enough of this to reconstruct what an agent did and why. I expect AI forensics to become a real specialization.

My experience with AI

I use AI every day. It helps me write and review code, draft documentation, translate content, and get a first understanding of unfamiliar topics. It saves me hours every week.

It also fails in ways that matter for our field. It confidently invents structure offsets, mixes up artifact versions, and suggests fields that do not exist. In many situations, I have to verify some claims against the source, the documentation, or a well-known tool. There is also a rule I do not break: real case data does not go to a model I do not control.

Developing DFIR tools

This is where AI changed my routine the most. In the past few months, I used it heavily while working on my new and refactoring my old projects. Work that would have taken me weeks or even months took days.

Tips and what I learned during this process:

What will happen to us?

I do not believe DFIR professionals will be fully replaced, at least not for some years. Someone still needs to understand the business from a human perspective, decide what matters, sign reports, and defend the findings in front of a client, a stakeholder, or a court.

What will change is what our work looks like: less time parsing and building timelines, more time validating, correlating, and making decisions. The professional who uses AI well will replace the one who does not and this is inevitable. The "downside" is that analysts may be overloaded: while a person could comfortably work on one or two cases per week, we can now work on many more; this may raise managers' expectations of their technical staff.

My biggest concern is with those starting now. Juniors learn by doing the repetitive work, such as parsing artifacts by hand, building timelines, reading assembly. If AI does all of that, how will the next generation build the knowledge needed to notice when the AI is wrong?