The defensive vendors

The defensive side has an incumbent the offensive side does not, and nobody has noticed it. It has been trading since 2005, it employs 201–500 people, and it sells the exact product a specialist would launch.

#Loginsoft is the price floor

Loginsoft markets "Security Data for AI Training" — "curated, labeled, and synthetic cybersecurity datasets" spanning "exploit detection, threat hunting, cloud security, and secure code review," with "expert labeling and ground-truth validation" (Loginsoft). A parallel "AI Model Validation" line builds custom "evaluation datasets that reflect real operating conditions, including edge cases and adversarial prompts," with "human-in-the-loop review" — expert scoring of model outputs for correctness and security relevance (Loginsoft).

That is expert labelling plus ground-truth validation plus custom eval sets. It is the product.

Company facts, all from a third-party directory rather than the company [WEAK]: founded 2005, HQ Chantilly, Virginia, development centre Hyderabad, 201–500 employees, billing $25–$49/hour, $25,000 minimum project, clients listed as HYAS, Scripps Networks, ThreatConnect and Verizon Labs (Enosis).

The honest read is that Loginsoft is a services firm extending into data, not a data company. No pricing, dataset volumes, customer names or scale figures appear on its own site — the rate is a directory's claim, not a rate card. It has no venture capital, no benchmark, no publication record and no lab logos. But it is selling, it is cheap, and a buyer running procurement will find it. Price against it deliberately or explain why the comparison does not hold.

#No AI SOC startup buys expert data

This is the most consistent finding across the defensive vendor set, and it survived checking each company individually. AI SOC startups treat expert knowledge as something to capture from customers in-product, never something to buy.

Sources: SiliconANGLE on Exaforce; Torq; Calcalist on Legion; Built In SF on Prophet; SecurityWeek on Conifers; ReliaQuest; SiliconANGLE on Cribl/Radiant.

Several hundred million dollars of venture capital in this column, and not one dollar of it disclosed as spent on external expert data.

#Legion is the closest analogue and the clearest anti-model

Legion Security built its entire thesis on the proposition a specialist would build on: analyst behaviour is the valuable artefact. Its own marketing says so — "real analyst behavior holds more value than any off-the-shelf playbook or pre-trained model" (Legion).

Its mechanism is a browser extension that "observes how analysts interact with tools like Chrome, Edge, or Island, and learns their decision-making patterns." That is a capture device for exactly the trajectory data a defensive expert-data company would pay practitioners to produce.

And then Legion does the opposite thing with it. It sells the capture mechanism to the customer, and the resulting data stays inside that customer's tenant. Nothing aggregates. There is no corpus at the end of it, because there is no company-level pool for one to form in.

Microsoft confirms the pattern from the other direction. Its production evaluation of the Defender Threat Detection Agent graded alerts using 1,088 alert-level grades from 208 customer organisations (arXiv 2605.20896). It bought the labels from customers, in kind, by running inside their tenants. Nobody in this column pays for labels in cash.

#The big vendors: telemetry they will not sell

This is a different competitive fact from having no data at all, and conflating the two produces a badly wrong map.

CrowdStrike has the data, publishes benchmarks, and will not sell the telemetry. It co-published CyberSOCEval with Meta, contributing real malware detonations from Falcon Sandbox (CrowdStrike). Its accompanying blog argues that meaningful evaluation requires "real telemetry from actual intrusions (not synthetic data)" and access to "trillions of events each day across a global customer base," and criticises public benchmarks for agenda capture, saturation and contamination (CrowdStrike).

That is a company staking out the position that only it can evaluate defensive AI properly. It is the mirror image of Irregular's contamination moat — hold the substrate rather than the measurement — and it is simultaneously a competitive threat and the strongest available validation of the thesis, and it is not wrong about SOC triage specifically — AuditBench's collapse from F1 1.00 on lab data to 0.25 on real OpTC data is a result, not a talking point (arXiv 2606.10281).

Meta publishes benchmarks and sells no data; CyberSOCEval is its move into the defensive half and it needed CrowdStrike to supply the substrate. Microsoft is the most active publisher of open defensive benchmarks — ExCyTIn-Bench and CTI-REALM, both open source, the latter headed for the UK AISI's Inspect repository — and discloses no annotation methodology or cost for either. Google/Mandiant publishes M-Trends and shipped Sec-Gemini v1, and no defensive benchmark or dataset could be established.

The structural point sits underneath all of them: Palo Alto Networks, SentinelOne and Microsoft all withdrew from the 2026 MITRE ATT&CK Evaluations, citing resource reallocation (Cybersecurity News). The industry's flagship independent evaluation is losing its largest participants at exactly the moment vendors are publishing their own benchmarks. Independent evaluation capacity is being vacated, and ten funded AI SOC vendors still need a way to say they are better than each other.

#What the map actually shows

Three columns, and only one of them is a competitor.

Selling the product: Loginsoft, cheaply, without venture backing or a public reference customer. That is the whole of the direct competition.

Sitting on data they will not sell: CrowdStrike, ReliaQuest, Microsoft, every AI SOC startup with customer tenants. These are not competitors for the buyer; they are competitors for the argument that external data is needed at all.

Capturing expertise without pooling it: Legion, Prophet, Tracebit. Validated premise, foreclosed model.

The absence worth naming: no venture-funded specialist exists on the defensive side. The only party visibly intermediating this work is a generalist — Mercor recruiting blue-team practitioners at $85–95/hr for an unnamed "cutting-edge AI research lab," which The generalists in cyber prices out in full. Defensive security and Defensive supply work through why the sub-tasks split the way they do — detection engineering and malware reverse engineering are fully public-buildable and demonstrably unsaturated; SOC triage is inseparable from customer telemetry and the incumbents are right about that.