Closing the Asymmetric Attack Surface: Corma Secures $125M to Industrialize Defensive Cyber Foundation Models

Corma achieves a $1B valuation to address the 85% to 12% LLM offense-defense detection disparity through purpose-built cybersecurity models.

Published: 2026.11.17

Visit Official Site
LLM Offensive vs Defensive Gap
85% vs 12%

Malware installation success versus self-detection rate

Investigation Latency Reduction
100x Faster

Reported breach response workflow acceleration

Capitalization and Post-Money Valuati…
$125M / $1B

Series A round led by NEA with Coatue and Sequoia

Executive Takeaway ·Frontier general-purpose LLMs fail at enterprise threat triage due to structural telemetry gaps, making domain-trained defensive models necessary for sub-…

Addressing the 85-to-12 Ratio in Model-Driven Cyber Operations

Enterprise security operations centers (SOCs) face a widening asymmetry in generative artificial intelligence. While commercial large language models synthesize exploit payloads, construct polymorphic evasion wrappers, and enumerate network perimeters with high consistency, their defensive remediation capabilities remain fundamentally impaired. Empirical network simulation tests conducted by Corma show that general-purpose frontier models achieved an 85% success rate in deploying unauthorized malicious payloads across segmented enterprise testbeds. When tasked with auditing the exact environments they breached, the same models detected only 12% of those files.

This disparity stems from core model training objectives. General-purpose models prioritize natural language synthesis and generalized code completion over high-dimensional graph reasoning across noisy, fragmented enterprise telemetry. Defending a production environment requires reconstructing sparse indicators of compromise (IOCs) across disparate endpoint detection and response (EDR) agents, firewall drop tables, and identity provider session caches. General LLMs fail in these environments because they cannot correlate out-of-order logs across abbreviated time horizons.

The Offense-Defense Performance Asymmetry in Frontier LLMs

Success rates across offensive penetration testing versus defensive file discovery

Offensive Exploitation vs Defensive File Detection
85% Malicious File Infiltration 12%
78% Lateral Movement Synthesis 22%
91% Security Policy Bypass 16%

Corma secured $125 million in Series A funding at a $1 billion post-money valuation to rectify this structural imbalance. Led by New Enterprise Associates (NEA) with participation from Khosla Ventures, Coatue, and Sequoia Capital, the capital is directed toward training foundation models natively optimized for defensive triage and automated remediation. Rather than retrofitting standard completion models with system prompts, Corma trains specialized architectures on raw telemetry streams to compress investigation times by up to 100x.

Comparative Architecture and DefenseBench V5 Benchmark Profiles

General-purpose models like GPT-6 Astra and Claude Opus 5.5 operate over natural language tokens, which introduces significant tokenization overhead and context window degradation when processing raw JSON event streams. Corma’s foundation model ingests raw, semi-structured network telemetry directly into high-dimensional embedding spaces tuned to temporal state-machine analysis. This architectural deviation prevents the model from hallucinating false-positive causal chains across overlapping IP address leases or intermittent authentication retries.

Independent evaluation on DefenseBench V5—a standardized harness evaluating defensive reliability, tool calling accuracy, and end-to-end task completion latency—demonstrates that Corma outperforms horizontal frontier models across triage workloads. While generic frontier models incur token consumption penalties and latency spikes exceeding 120 seconds during alert correlation, specialized defensive architectures execute autonomous hypothesis testing in under two seconds.

DefenseBench V5 Benchmark Execution Latency

Time required to ingest multi-source telemetry and execute incident correlation

General Frontier LLM (GPT-6 / Opus 5.5) Corma Cyber Foundation Model
Multi-Source IOC Ingestion 70x Faster
General Frontier LLM (GPT-6 / Opus 5.5) 84s
Corma Cyber Foundation Model 1.2s
False-Positive Alert De-duplication 124x Faster
General Frontier LLM (GPT-6 / Opus 5.5) 112s
Corma Cyber Foundation Model 0.9s
Remediation Script Synthesis 18x Faster
General Frontier LLM (GPT-6 / Opus 5.5) 38s
Corma Cyber Foundation Model 2.1s

The performance profile directly impacts Tier-1 and Tier-2 SOC economics. General-purpose API inference requires extensive retrieval-augmented generation (RAG) scaffolding, vector database querying, and multi-agent coordination frameworks that drive processing costs upward. In high-throughput enterprise environments processing 100,000 security alerts per day, running horizontal frontier models creates unsustainable API token expenditures alongside prohibitive p95 latency queues.

Metric / Operational VectorCommercial General LLM (API-Driven)Open-Source Fine-Tuned LLaMA-70BCorma Domain-Specific Foundation Model
DefenseBench V5 Reliability Score64.2%51.8%94.6%
P95 Alert Correlation Latency45–90 seconds18–35 seconds0.8–1.5 seconds
Input Ingestion ModalitySerialized JSON / Re-tokenized textPrompt-engineered text tokensRaw multi-source telemetry tensors
Telemetry Ingestion Cost (100k events/day)~$1,800–$3,200 / day~$450 / day (Dedicated GPU compute)Enterprise SaaS license tier
Hallucinated Remediation Rate18.4%27.1%< 0.5%
Log Window ResiliencyRequires extensive continuous contextFails on truncated logsReconstructs sparse state from partial logs

Telemetry Ingestion Constraints and Truncated Log Windows

The primary technical bottleneck in automated incident investigation is not model reasoning capacity, but the structural degradation of enterprise log data. Financial pressures on enterprise infrastructure force infrastructure engineering teams to establish aggressive log retention policies. High-volume firewall egress logs, DNS request archives, and NetFlow tables are frequently truncated or routed to cold storage after 7 to 14 days to curb SIEM ingestion and indexing costs.

When advanced persistent threats (APTs) remain dormant for months before initiating exfiltration, autonomous investigation agents encounter massive telemetry voids. General-purpose models collapse under these conditions, attempting to fill data gaps with probabilistic hallucinations. Corma’s underlying architecture incorporates inductive biases designed specifically for temporal graph completion, allowing the engine to evaluate whether isolated process executions represent legitimate administrative tasks or lateral threat staging despite missing intervening network telemetry.

Autonomous Investigation Pipeline Under Sparse Telemetry Constraints

State-machine reconstruction across fragmented enterprise telemetry

01 · Step 01
Telemetry Ingestion

Ingests raw NetFlow, EDR events, and firewall drops without schema normalization

02 · Step 02
Graph State Reconstruction

Interpolates missing intermediate events caused by aggressive log retention windows

03 · Step 03
Autonomous Investigation

Executes automated forensic hypotheses and isolates validated attack paths

04 · Step 04
Actionable Orchestration

Generates deterministic remediation scripts and surfaces escalation alerts via Slack

To maintain operational guardrails, Corma splits response mechanics into fully autonomous triage and human-steered remediation. For recurring, low-entropy security alerts, the model isolates host processes, updates local endpoint rules, and verifies containment status independently. For complex network topology alterations or root credential revocations, the system packages forensic evidence into Slack operational channels. Administrators can interact with the engine conversationally to verify malware eradication, request secondary audits, or adjust remediation parameters prior to execution.

Deployment Trade-Offs and Enterprise SOC Readiness

Implementing domain-specific defensive models introduces distinct architectural trade-offs between centralized software-as-a-service (SaaS) execution and data residency boundaries. Enterprise CISOs must determine whether transmitting real-time network flow metadata and endpoint telemetry across external infrastructure satisfies compliance boundaries under frameworks such as SOC 2 Type II, FedRAMP, and ISO 27001.

Trade-Off Matrix: Specialized Defensive Model Adoption

Operational balance between automated speed and enterprise governance

Defensive Operational Advantages
  • + Sub-minute mean time to investigate (MTTI) across noisy alert streams
  • + Drastic reduction in Tier-1 analyst triage fatigue and false-positive churn
  • + Reconstruction of persistent security context across truncated log windows
Implementation and Governance Costs
  • – Telemetry egress exposure and cross-tenant data governance considerations
  • – Integration friction across legacy on-premises SIEM architectures
  • – High enterprise software licensing overhead displacing legacy SOAR tooling

Organizations operating modern security stacks should evaluate Corma against rigid operational parameters:

  • Immediate Deployment Indicators: Security operations centers triaging upwards of 50,000 unaggregated alerts daily, environments experiencing high Tier-1 analyst attrition, and hybrid-cloud infrastructures maintaining standardized API access across modern EDR tools.
  • Postponement Indicators: Highly air-gapped on-premises data centers with strict regulatory bans on external metadata egress, organizations lacking centralized telemetry pipelines, or environments where incident investigation requires physical hardware access.

Related Benchmark Reports

View All in Category →
Weekly Briefing

Weekly Tech & Business Data Briefing

Verified software analysis, practical gotchas, and essential supply-chain updates delivered weekly.

Unsubscribe with 1 click anytime. Zero spam.

* We may earn an affiliate commission from links in this report, at no extra cost to you and with zero impact on our benchmark data.