Techno Security West 2026: AI, the Dark Web, and Ghost Machines
Three days in San Diego: agentic AI investigations, ghost machines hiding inside compromised hosts, preserving AI prompts as evidence, and what the dark web actually is.
In early October I posted that I would be at Techno Security West in San Diego from the 20th through the 22nd. I was there all three days, and the trip paid off.
This was only my second Techno West, but I have attended and spoken at the original Techno Security conference in Myrtle Beach for many years, so I traveled with expectations. The room was the usual Techno mix: junior examiners finding their footing next to senior people who have been doing this since long before most of these cases existed, with law enforcement and the private sector sitting in the same sessions. I attended presentations in nearly every time slot, and when a slot did not offer anything compelling I used the time to work the room and catch up with old friends. That has always been the point of this conference.
An honest observation: a few sessions left me wanting more depth, and at least one was closer to a product pitch than a lecture. I still took something useful away from all of them, and more than one planted ideas for tools I want to build. Colleagues I know and trust who also attend Techno East say the Myrtle Beach edition runs a bit higher quality and draws at least 30 percent more attendees. I would put this crowd between 300 and 500 people, though attendance is a metric everyone at every conference argues about.
Agentic AI for OSINT investigations
Chet Hosmer and Serena opened the AI thread with a look at how agentic AI changes investigations. The shift they described is the one that matters: instead of answering prompts, these systems pursue explicit objectives. Give them a properly engineered objective statement and a multi-agent workflow, and they can run large-scale, evidence-driven OSINT investigations with automated validation and report generation at the end.
I have known Chet for years, and he knows his material cold. His walk through a framework for applying AI to OSINT work on real events was the clearest treatment I have seen of putting objective engineering ahead of prompt craft.
What the dark web actually is
Todd Shipley gave the entry-level talk on the dark web, and Todd is the right person to do it. He walked through the layers everyone gets wrong: the clear web, the deep web, and the dark web, which is mostly Tor. From there he explained how anonymity works on those networks, what they get used for, and what investigators can lawfully collect.
I know Todd well, and this is his home turf. For anyone who has never opened the Tor Browser, this was the sane introduction: no mystique, just routing, anonymity, and the practical realities of gathering evidence on a network built to resist attribution.
Ghost machines
Jeff Stanton’s talk was the sleeper hit of the week for me. Ghost machines are virtual machines running inside a host the attacker already owns. A compromised box quietly runs a VM that holds the attacker’s tooling, can run AI agents that automate exploitation, and sits outside the reach of host-centric detection.
The lesson for defenders: you have to hunt for the VMs themselves, watch network telemetry, and know what artifacts these setups leave behind on macOS and Windows. Stanton described it as an introduction, but it was an excellent one, and it lit up the part of my brain that designs forensic tools. My notebook for that session filled up fast.
Preserving AI prompts as evidence
The most practical AI session of the conference came from the two Digital Mountain presenters: Kurt Ekensten, a project manager and forensic examiner, and Michael Mansuy, the VP of operations. They covered the current state of generative and agentic AI, what it means for legal and infrastructure teams, and concrete guidance for preserving and reviewing AI prompts across cloud, mobile, and endpoint sources.
Their central complaint landed exactly where my own work landed: there is essentially no tooling for reconstructing local AI conversation chains from the machine itself. That is the gap documented in my white paper Forensic Artifacts Left by AI: Tool Capability Needs, and the gap the AI Extract and Analysis Workbench is built to close.
Their approach reconstructs chat sessions from the cloud provider’s data when it exists. That is a viable path when the cloud copy is available and collectable. It is not the only path, and it is not available in every case. The local artifacts are where AI Extract works.
Verifying what AI tells you
Josh and Mobasi co-founder Tyler made the case that AI can surface investigative leads quickly, but speed is worthless if the examiner cannot trace, cite, and verify the output against trusted forensic sources. Findings have to be defensible and reproducible, or they are not findings.
The deck was nearly pure product pitch. It was also one of the most complete agentic AI forensics platforms I have seen. Mobasi’s agents operate on exports from the mainstream tools, ingesting reports and images from EnCase, FTK, and Cellebrite, and they document their work along the way. Whatever you think of AI in the lab, this is where the industry is moving.
Evidence from the dark web, the hard way
Todd Shipley came back for a second, deeper session on collecting and interpreting evidence from Tor and the other public anonymity networks. His main point deserves repeating: screenshots are not evidence. The artifacts that matter live in browser communications, web-server responses, and site source code. If you are building dark web cases on screenshots, you are building them on sand.
This session did what I hoped it would: it filled another page with ideas for tools that close gaps in live OSINT collection.
The theme that tied it together
Look at the sessions I picked and one through-line appears: AI has fully arrived in digital forensics. Attackers run AI agents inside ghost machines. Investigators use agentic AI to scale OSINT. Examiners now have to preserve AI prompts as evidence and verify what AI tools tell them before it goes anywhere near a report. The community is still working out how to handle all of it in a way that survives cross-examination, which is exactly why those three days in San Diego mattered.
If we crossed paths at Techno West, I am glad we did. If you were not there, keep an eye on Techno East in Myrtle Beach. After this year, I would like to compare the two editions for myself.