For Atlas, this made it a good opportunity to put its approach to the test.
Why a public benchmark matters
Claims about AI-powered digital forensics are easy to make. Demonstrating that the results are actually reliable is a different matter. Real incident investigations are confidential, and their evidence cannot simply be published for others to inspect.
Public forensic challenges offer an alternative. The evidence is available to everyone, the original questions are documented, and the expected findings can be compared against an official solution. Anyone can repeat the investigation and verify the results.
That is why we chose Rhino Hunt.
The case
Originally developed as a hands-on exercise for the DFRWS 2005 conference, Rhino Hunt is now part of the NIST CFReDS archive.
The scenario is fictional. A law in New Orleans prohibits collecting pictures of rhinos, and a computer seized from a university lab is suspected of containing evidence of such a collection. The original hard drive is missing. What remains is an image of a USB drive and three network traces captured by the university's network administrator.
The challenge consists of seven questions. Finding the answers requires examining both the USB image and the network traffic. Neither source tells the complete story on its own. The relevant information is spread across file remnants, network sessions and the relationships between them.
This makes Rhino Hunt particularly interesting for Atlas, which is designed to analyse different evidence sources and correlate their findings.
How we tested Atlas
For this run, Atlas used Qwen3.8-27B-FP8 as its language model.
Atlas received the evidence along with a short case brief outlining the questions to be answered and the areas worth investigating. The brief is published alongside the report, allowing anyone to see exactly what information was provided before the analysis began.
Beyond that, Atlas had to work with the evidence itself. Findings needed to be supported by traceable evidence, and anything that could not be established had to be identified as an open question rather than filled in with assumptions.
Once the investigation was complete, we compared the resulting report with the official DFRWS answer key.
What Atlas found
Atlas reconstructed most of the relevant activity across the USB image and the three network traces.
It recovered the account name and password, gnome / gnome123, directly from a plaintext FTP session. It also identified the relevant file transfers, including rhino1 and rhino3 as well as an encrypted ZIP archive sent over FTP. The HTTP traces contained further transfers involving rhino4, rhino5 and an executable. Each transfer was documented with its corresponding timestamp and host information.
On the USB drive, Atlas recovered a diary from the reformatted filesystem. It describes how the original hard drive was thrown into the Mississippi, why the USB drive was reformatted to conceal the pictures, and how the account credentials were obtained from Jeremy.
The only files still visible on the reformatted drive were two gumbo recipes. Atlas identified them as decoys and also used them, together with the diary, to derive candidate passwords for the hidden data.
Most importantly, Atlas connected the two sides of the investigation. The account mentioned in the diary was the same account used for the network transfers.
The analysis also uncovered something beyond the official answer key. In the network traces, Atlas found a Hotmail exchange originating from the same machine during the FTP upload. The correspondence provides a name for the person behind the activity.
How the results compare
Atlas answered six of the seven challenge questions correctly. One answer differed from the official solution.
For the question concerning the connection between the USB drive and the network activity, the answer key points to a picture found in both sources. Atlas instead established the connection through the shared account. Given that the answer key explicitly accepts “at least one thing” as a valid connection, this is a different route to establishing the relationship rather than an unsupported conclusion.
The question about what could be recovered from the USB drive was only partially answered. Atlas carved the diary and six JPEG files, including the two images used as steganographic carriers. It also derived the correct password candidates from the diary and the decoy files.
Question by question: Atlas and the official answer key
| Question | Answer key | Atlas | Result |
|---|---|---|---|
| Who gave the accused the account? | Jeremy (from the diary) | Jeremy, from the carved diary | Correct |
| Username and password? | gnome / gnome123 | gnome / gnome123, from the FTP session | Correct |
| Which file transfers are relevant? | rhino1, rhino3 and contraband.zip over FTP; rhino4, rhino5 and rhino.exe over HTTP | All six, each with time and host | Correct |
| What happened to the hard drive? | Thrown into the Mississippi | Thrown into the Mississippi, from the diary | Correct |
| What happened to the USB key? | Reformatted | Reformatted, with two gumbo recipes left as decoys | Correct |
| What can be recovered from the USB image? | The diary, the recipes and carvable pictures, two of them hidden with JPHide | Diary and six JPEGs carved, password candidates derived; hidden payloads not extracted | Partly |
| What connects the USB key and the network traces? | rhino2.jpg on the key matches the picture in the ZIP | The gnome account: named in the diary, used for the transfers | Correct |
Knowing where to stop
Not everything could be recovered automatically.
Two pictures were hidden using JPHide, a steganography tool that requires JPHide for Windows to extract their contents. The tool is considerably older than the challenge itself and was not available in Atlas's toolset.
Rather than attempting to interpret the files without the necessary tooling, Atlas documented the limitation in its report. It identified the missing information, explained why it could not proceed and specified which tool would be needed to complete that part of the investigation.
The remaining open points are documented in the report as well.
For a forensic tool, knowing what cannot be established is just as important as reporting what can. An incomplete finding presented as fact is worse than an explicitly documented limitation.
From a public challenge to real investigations
Rhino Hunt is a relatively small case, but the underlying work is familiar to anyone involved in incident response. Evidence comes from different sources, traces have been deliberately removed or concealed, and individual findings only become meaningful once they are connected.
This is the kind of work Atlas is built to handle. It extracts artifacts, runs established forensic tools, correlates evidence across sources and produces a structured first assessment. Every finding links back to the evidence and the tool call that produced it.
The goal is not to replace the examiner. It is to move the starting point of an investigation. Instead of manually collecting and sorting every artifact, the examiner can begin by reviewing an existing body of findings, checking the underlying evidence and focusing on the questions that still require hands-on analysis.
Try the investigation yourself
The Rhino Hunt investigation is included in the Atlas Community Edition as a demo. You can find it under Dashboard → RHINO-HUNT, together with the original case brief and the complete analysis report.
The underlying evidence is publicly available through the NIST CFReDS archive. You can run the challenge yourself and compare the results with Atlas's findings.