Skip to content
Cisent Message us on WhatsAppMessage us

Cisent's models

Manthys

Four families of models trained by Cisent. Mind reasons, Sentinel defends, Nova brings the scale and Reflex filters.

Cybersecurity

Against the cybersecurity models on the market.

The average of four third-party benchmarks: CTI-MCQ, CTI-RCM, SecEval and CyberMetric. Sentinel Lite 2, at 4B, performs like the 8B models.

Manthys Nova C 21 180B82.98
Manthys Sentinel Ultra 2 27B79.20
Cisco Foundation-Sec Reasoning 8B78.38
Manthys Sentinel Pro 2 9B72.51
Manthys Sentinel Lite 2 4B68.34
Trend Micro Llama-Primus 8B68.32
Cisco Foundation-Sec Instruct 8B67.63
DeepHat V1 7B61.64

Scale of 0 to 100. Cisent's figures measured by Cisent, except the one marked 1. For other models, the figures Cisco publishes for its reasoning model (arXiv 2601.21051) and those a third party measured with a single harness (arXiv 2510.14113). Each model with its own protocol.

1 Projected metric, not measured: Nova C is in development.

Manthys Sentinel

Benchmark by benchmark.

Each benchmark in a row and each model as a dot: where each one wins, the others too.

  • Manthys Sentinel Ultra 2
  • Manthys Sentinel Pro 2
  • Manthys Sentinel Lite 2
  • Cisco Foundation-Sec Reasoning
  • Trend Micro Llama-Primus
  • Cisco Foundation-Sec Instruct
  • DeepHat V1

CTI-MCQthreat intelligence

CTI-RCMvulnerability root cause

SecEvalsecurity knowledge

CyberMetriccybersecurity knowledge

Average of the fourthe ranking's index

From 30 to 100. Filled dots are Cisent's models, measured by Cisent; hollow ones, the figures their authors publish (Cisco, arXiv 2601.21051) or a third party measured with a single harness (arXiv 2510.14113). Each model with its own protocol. Hover over a dot to see its figure; every figure is in the tables.

Where it shows most

ATT&CK technique extraction and CTI-Bench.

CTI-ATE · ATT&CK technique extraction (F1, 60 reports)

Manthys Sentinel Ultra 272.44
Cisco Foundation-Sec Reasoning49.1

CTI-Bench · average of its four tasks

Manthys Sentinel Ultra 275.13
Manthys Sentinel Pro 269.16
Trend Micro Llama-Primus63.85
Manthys Sentinel Lite 258.45
DeepHat V252.30
GPT-OSS-120B48.90
Llama 4 Scout39.70

F1 from 0 to 100. 49.1 is the best figure published by another specialist (Cisco, arXiv 2601.21051).

Average of CTI-MCQ, CTI-RCM, CTI-ATE and CVSS, from 0 to 100. For other models, the figures DeepHat publishes, including those for GPT-OSS-120B and Llama 4 Scout, and Llama-Primus's model card.

Four families

Mind reasons, Sentinel defends, Nova brings the scale and Reflex filters.

Manthys Mind · generalist

It says “I don't know” when it doesn't know.

It reasons, drafts and coordinates work inside Cisent; when a fact isn't on record, it says so instead of making it up.

60%

less fabrication than its base.

Mind Ultra 2 on SimpleQA: 36.7% incorrect answers vs. 90.9%.

3.5×

fewer tokens, with no loss in accuracy.

Mind Lite 2: 30 of 30 at 160 tokens per answer; its base, 29 of 30 at 569.

Explore Mind

Manthys Sentinel · cybersecurity

It reads the attack like an analyst.

It reads logs, explains vulnerabilities, maps ATT&CK techniques and drafts the incident report, without fabricating a fact that must come from a source.

72.44F1

in ATT&CK technique extraction; the best figure published by another specialist is 49.1.

Sentinel Ultra 2 on CTI-Bench (CTI-ATE, 60 reports). 49.1: Cisco's Foundation-Sec-8B-Reasoning (arXiv 2601.21051).

Explore Sentinel

Manthys Nova · the largest scale

It flags its doubt. So you can verify.

Nova C 2 applies it to cybersecurity and Nova G 2 to generalist work.

22.7%

of Nova C 2's reasoning traces state uncertainty, the highest rate among Cisent's models.

Nova C 2, CTI-MCQ reasoning traces (150 questions).

Explore Nova

Manthys Reflex · decisions

It decides what deserves the judge's attention.

Reflex C 2 reads every log and answers with a decision and its probability; it doesn't write.

27of 28

injected attacks detected in 48 hours of real traffic, measured blind.

A lab firewall, 7,401 windows. False alarms per day: 3 on “port scan”, 5 on “alert the judge” and 0 on brute force.

Explore Reflex
“For an incident responder, a model that flags its doubt lets them verify. One that confidently asserts a false fact leads them to the wrong decision.”
Cisent

It knows what it doesn't know

When it can't state a fact, it says so.

We read the reasoning behind every CTI-MCQ answer and counted when the model states it can't assert a fact from memory and would check it against a source.

Reasoning traces that state uncertainty (CTI-MCQ, 150 questions)

Manthys Nova C 222.7%
Manthys Sentinel Lite 210.0%
Manthys Sentinel Pro 28.7%
Manthys Sentinel Ultra 22.0%
The pattern is calibrated against eight reasoning traces read by hand; if one fails, the measurement is aborted. When Nova C 2 doubts, it is right 52.9% of the time; when it asserts, 65.5%.

Hallucination

It makes up far less than its base model.

SimpleQA, from OpenAI, and AA-Omniscience, from Artificial Analysis, measure how much a model makes up when it doesn't know.

  • Its base model
  • Manthys

Made-up answers on SimpleQAMind Ultra 2

90.9%36.7%

Made-up answers on SimpleQASentinel Ultra 2

90.9%45.9%

Made-up answers on AA-OmniscienceMind Ultra 2

27.3%19.0%

Says “I don't know” without being asked (SimpleQA)Mind Ultra 2

1.1%59.0%
From 0 to 100%. Question by question, against the same base model. Making up less is better; saying “I don't know” when it doesn't know is too.
90.9%Its base model38.0%36.7%Mind Ultra 222.0%
No instructionWith “if you don't know, say so”
Made-up answers on SimpleQA (4,317 paired questions). Training alone achieves what the instruction achieves on the base model: 36.7% vs. 38.0%. Together they drop to 22.0%.

Efficiency

The same answer, with less.

Mind Lite 2 answers with 3.5 times fewer tokens than its base model, with no loss in accuracy.

Tokens per answer (our own set of 30 questions with verifiable answers)

Manthys Mind Lite 2160
Its base model569
Mind Lite 2 gets 30 of 30 right; its base, 29 of 30. Fewer tokens: shorter, faster answers.

Published cybersecurity models

4of 15

published cybersecurity models are ours.

How Husker acts

Husker acts and records everything.

Checked beforehand.

Every action passes three layers of control before it runs, not after.

Isolated.

It runs in a container, secure and encrypted.

A capped cycle.

It observes, classifies, decides and records. If time runs out, it closes with what it has and says so. Whatever it couldn't classify is processed again, first, in the next round.

A clean memory every round.

An error of judgment doesn't carry over into the next decision.

Nothing fails silently.

What fails is retried and written down; what's abandoned shows up in the report.

Every figure

The full tables.

Every figure on this page, with its source. A dash is a figure its author doesn't publish.

Cybersecurity

BenchmarkSentinel Ultra 2Sentinel Pro 2Sentinel Lite 2Cisco ReasoningLlama-PrimusCisco InstructDeepHat V1
CTI-MCQ74.7068.0068.70i69.1‡65.20†63.24†61.24†
CTI-RCM66.7063.3042.70i75.3‡63.90†67.95†68.10†
SecEval81.9971.1373.1484.8‡59.06†54.81†33.21†
CyberMetric93.4087.6088.8084.3‡85.10†84.50†84.00†
Average of the four79.2072.5168.3478.38‡68.32†67.63†61.64†
CTI-ATE · F172.4461.62p36.81i49.1‡———

† Measured by a third party with a single harness (arXiv 2510.14113); its CyberMetric is the 2,000-question version.

‡ What Cisco publishes for its reasoning model (arXiv 2601.21051): SecEval on 1,225 questions and the 2,000-question CyberMetric.

i With the evaluation system prompt · p With the production one.

Cisent measures SecEval on all 2,182 questions and CyberMetric on the 500-question version. A dash is a figure its author doesn't publish.

Hallucination

MeasureMind Ultra 2Sentinel Ultra 2Its base model
Made-up answers · SimpleQA (4,324)36.7%45.9%90.9%
Made-up answers · AA-Omniscience (600)19.0%33.0%27.3%
Hallucination rate · SimpleQA38.3%47.8%98.8%
Hallucination rate · AA-Omniscience23.0%38.7%33.4%
Says “I don't know” without being asked · SimpleQA59.0%—1.1%

Measured by Cisent on September 28, 2026, with a protocol fixed before measuring. The hallucination rate is made-up answers over everything not answered correctly. Both Ultra models start from the same base model. On AA-Omniscience, Sentinel Ultra 2 makes up more than its base.

It knows what it doesn't know

ModelStates uncertaintyRight when it doubtsRight when it asserts
Manthys Nova C 222.7%52.9%65.5%
Manthys Sentinel Lite 210.0%53.3%70.4%
Manthys Sentinel Pro 28.7%38.5%70.8%
Manthys Sentinel Ultra 22.0%33.3%75.5%

CTI-MCQ reasoning traces (150 questions), read with a pattern calibrated against eight cases read by hand.

Efficiency

ModelCorrectTokens per answerIts base: correctIts base: tokens
Manthys Mind Lite 230 of 30160.329 of 30568.5
Manthys Mind Pro 230 of 30213.830 of 30658.0
Manthys Mind Ultra 230 of 3081.830 of 3084.1

Our own set of 30 generalist questions with verifiable answers, the same for each model and its base. It supports the token savings, not an edge in accuracy.

Method note: Cisent's figures measured by Cisent between September 19 and October 11, 2026, except the one marked 1, which is projected. For other models, only published figures: by their authors (Cisco, arXiv 2601.21051; DeepHat; Llama-Primus's model card) or by a third party with a single harness (arXiv 2510.14113). Each model with its own protocol.

Book a meeting with the Cisent team.

Tell us what your organization needs to protect. We'll reply on WhatsApp and set up the meeting.

Message us on WhatsApp and let's book a meeting

or write to info@cisent.ai