M&NTIS Platform v26.10: What’s New

We’re pleased to announce M&NTIS Platform v26.10, a release focused on AI, from both sides of the fight. On the defence side, a new scenario, Delusia, puts AI agents wired into a company’s internal tooling at the heart of the investigation. On the attack side, the new mantis bench command lets an LLM-driven attacker loose in a lab to challenge its detection. The release also makes a lab easier to read, and makes labs fail early and say why.

⭐ Key Highlights

Delusia: When AI Agents Go Off Script

AI agents now read tickets, open merge requests and run commands inside companies’ tooling. Delusia brings that new attack surface into a lab.

The targeted organization is a software vendor in the technology sector, where protecting customer data, source code and production secrets is critical. The company recently deployed AI agents to handle customer support and routine development tasks. Unusual activity has since been reported around these agents, starting with notes on customer tickets that no technician remembers writing. The investigation must determine the origin of this activity and assess its scope, before it leads to data leaks, loss of customer trust and regulatory penalties.

A detection signature for the first phase is available for Kibana, Graylog and Splunk. The alert it triggers is the starting point of the investigation for the SOC and CERT analysts in training.

M&NTIS-Bench: An LLM Attacker to Challenge Your Defences

Attackers are starting to rely on AI to speed up their intrusions. M&NTIS-Bench lets a team find out how its detection holds up against such an attacker, in a lab, before facing one for real.

The new mantis bench command of the M&NTIS CLI gives an LLM an initial access in a running lab and lets it attack autonomously, from discovery to privilege escalation and lateral movement. When the model stops, the command scores what the lab’s detection caught, straight from the lab’s attack report and the alerts raised by its SIEM:

• the detection rate: the share of the attacker’s actions that raised a correlated alert on the same host;
• the time to detect, and how many actions each collector caught.

Use it to qualify a SIEM configuration or a ruleset against an attacker that has not read your detection rules, or to compare how your defence copes with attackers driven by different models. The model can be Claude, Qwen, or any OpenAI-compatible endpoint, including a self-hosted model when the commands run in the lab must not leave your premises.

📋 Other Release Changes

A Lab That Is Easier to Read

• The Topology tab is split into a Network view, the graph of the lab, and an inventory view listing each machine with its addresses, networks and installed software side by side.
• A lab infrastructure section lists the platform nodes kept off the graph, such as the provisioning node, and explains the traces they leave.
• The lab content start time is shown in that section and in the lab timeline, and a notification announces it. Anything before that time is lab preparation, not the scenario: analysts no longer chase provisioning traces as if they were the attacker.

Labs That Fail Early and Say Why

A lab whose result could not be trusted now stops, with a notification naming the cause, rather than reporting a misleading result at the end.

• A detection rule that cannot be pushed to the SIEM now stops the lab. Until now, the attack was reported as undetected: a false negative presented as a result.
• Auditbeat moves to 9.2.0 and must stay up. On Ubuntu 24.04 machines, the previous version stopped a few seconds after starting, and its events were silently missing. A lab whose Logstash stops now fails too.
• A scenario’s missing required parameters, and a topology using a node name the lab adds itself, are now refused before any resource is reserved, instead of failing up to an hour later.
• While cloud capacity is short, the user is kept informed every 15 minutes, the lab’s maximum duration only starts once its machines are reserved, and a lab that gets no capacity in time ends Cancelled rather than in error.

Detection and Integrations

• The Suricata probe loads the Emerging Threats Open ruleset next to the M&NTIS rules, with the categories that mostly report legitimate lab activity muted.
• Sekoia.io: every hosting region is supported, Suricata EVE logs are forwarded to a dedicated intake, and the connector now only deletes the objects a lab created, at the end of the lab.
• A log collector setting can now concern some operating systems only, such as one onboarding package per OS family, and file settings can be uploaded from the configuration dialog.

Interface

• Notifications name the machines a provisioning step, a user activity or a log collector targets.
• The assessment (quiz) is off by default and flagged beta. Answers are now compared regardless of case and surrounding spaces.
• Guacamole RDP connections use the keyboard layout of the guest, which fixes AZERTY typing on Kali.
• The SOC/CERT training guide is linked from the lab creation flow, and trial users are pointed to the free trial quick start.

Scenarios and Content

• Linux machines of a lab now reach the Internet through the lab’s proxy, like Windows machines already did, for command-line tools, APT and Firefox.
• The GOAD labs now run on the same addresses as the upstream GOAD project, so its walkthroughs apply as is.

For API and CLI Users

• The lab returns a new lab_content_start_timestamp field: the end of lab preparation, when the scenario content starts.
• A lab created without a scenario’s required parameters is refused with an HTTP 400 error naming them.

M&NTIS v26.10 brings AI to both sides of the exercise: AI agents become an attack surface a team can investigate, and an LLM becomes an attacker a defence can be measured against.