Cosmipher Freedom V1.1
Security-specialized intelligence for researchers, security teams and enterprises.
Cosmipher Freedom V1.1 is a 602B logical-parameter, multimodal mixture-of-experts model built in India for cybersecurity research, authorized testing, and enterprise defense and offense. Its capability atlas spans 51 security domains and 431 scenarios, connecting AI-agent security with application security, cloud infrastructure, incident response, cryptography and industry-specific systems.
Bring your security use case to Cosmipher, whether you are exploring a research question, evaluating a private endpoint or building a workflow around your team's existing tools.
Request access
Individual researchers and organizations must submit the Cosmipher contact form to request model access. The form has six quick questions about you and your team and takes about a minute.
Explore Cosmipher · Official Freedom V1.1 page
Breadth across the cybersecurity discipline
Freedom's capability atlas covers 391 domain scenarios, 20 self-contained showcase cases and 20 historical CVE challenges. The six areas below organize that breadth around the work security teams actually do.
AI and agent security
9 domains · 74 domain scenarios
Examine how models, agents, tools and retrieved information interact. The atlas covers goal hijacking, tool misuse, privilege boundaries, prompt injection, poisoned retrieval, MCP integrations, multi-agent delegation and AI artifact provenance.
- Agents and delegation — goal hijack, memory and context poisoning, approval bypass, privilege propagation and compromised-agent isolation.
- Prompt injection and retrieval — retrieved-document overrides, tool-result laundering, cross-tenant vector leakage, citation laundering and semantic-cache isolation.
- MCP and connectors — tool-schema trust, OAuth scopes, tool-output injection, server identity and connector exfiltration risk.
- Models and AI supply chains — dataset poisoning, model-weight integrity, adapter provenance, tokenizer changes, quantization correctness and inference artifact parity.
- Evaluation and operations — benchmark contamination, judge disagreement, tool-trace fidelity, model-card integrity, inference authentication and prompt-log privacy.
Application security
9 domains · 72 domain scenarios
Explore the trust boundaries inside web applications, APIs, mobile clients, native code and smart contracts. Coverage extends from authentication and business logic to software dependencies and secure development practices.
- Authentication and sessions — OAuth redirect policies, JWT validation, reset tokens, MFA recovery, CSRF and session revocation.
- Web and API weaknesses — BOLA/IDOR, function-level authorization, SQL injection, stored XSS, SSRF, path traversal and unsafe deserialization.
- Secure engineering — security-focused PR review, threat modeling, SAST/DAST triage, property-based testing and secrets in delivery pipelines.
- Mobile and native code — exported components, credential storage, certificate validation, memory safety, race conditions and fuzz-harness design.
- Supply chain and Web3 — dependency confusion, SBOM reconciliation, build attestations, Solidity access control, reentrancy and bridge trust.
Cloud and infrastructure
11 domains · 82 domain scenarios
Follow identities, permissions, workloads and network boundaries across the major clouds and the systems that support them.
- AWS — IAM permission graphs, cross-account trust, S3 exposure, CloudTrail timelines, KMS governance and VPC reachability.
- Azure and GCP — Entra conditional access, RBAC scope, managed identities, service-account impersonation, workload federation and cloud audit evidence.
- Kubernetes and containers — RBAC escalation, admission controls, secrets, workload trust, image provenance, namespace isolation and multi-tenant boundaries.
- Operating systems and endpoints — Linux capabilities, systemd persistence, auditd, Kerberos, Active Directory, GPOs, browser-extension permissions and updater trust.
- Networks and wireless — segmentation, firewall rules, DNS tunneling, TLS downgrade analysis, VPN policy, Wi-Fi, BLE, LoRaWAN and telecommunications scenarios.
Detection and response
9 domains · 69 domain scenarios
Connect technical signals with investigation, detection engineering, vulnerability prioritization and defensible incident decisions.
- SOC and detection engineering — Sigma creation, Splunk SPL and KQL translation, signal correlation, severity calibration and detection regression suites.
- Threat hunting — hypothesis-driven hunts, beaconing analysis, living-off-the-land behaviour, data staging and cross-system timeline alignment.
- Incident response — ransomware and cloud-key scenarios, identity emergencies, recovery integrity, executive briefs and post-incident review.
- Malware, reversing and forensics — behaviour reports, sandbox summaries, YARA quality, persistence, binary analysis, evidence preservation and log-tampering analysis.
- Vulnerability and exposure work — CWE root causes, patch regression, reachability-based prioritization, passive DNS, certificate transparency and authorized purple-team exercises.
Data, identity and cryptography
6 domains · 42 domain scenarios
Assess who can act, which data they can reach, and how identity and cryptographic controls protect those boundaries.
- Data protection — PII flows, database row-level security, DLP, backup immutability, deletion semantics and data lineage.
- Applied cryptography — AEAD nonce reuse, JWT/JWS key confusion, password storage, TLS validation, replay and key lifecycle governance.
- PKI and zero trust — certificate-authority incidents, mutual TLS, certificate issuance, workload identity and HSM isolation.
- Post-quantum migration — crypto-agility inventories, hybrid key establishment, long-lived data exposure and certificate migration.
- Fraud and phishing defense — account-takeover signals, transaction graphs, synthetic identities, business email compromise, OAuth consent phishing and DMARC migration.
Industry, people and research
7 domains · 52 domain scenarios
Extend security reasoning into physical systems, industrial operations, governance and the human decisions surrounding technical controls.
- OT and IoT — PLC zoning, Modbus anomalies, vendor access, firmware maintenance, embedded interfaces and connected-device boundaries.
- Cyber-physical systems — badge access, camera and device exposure, tamper-evident evidence and operational controls.
- Governance and risk — NIST CSF mapping, ISO 27001 gaps, SOC 2 evidence, privacy impact, vendor risk and policy-to-control traceability.
- Human factors — phishing awareness, approval fatigue, incident communication, security usability and calibrated trust in AI.
- Research discipline — competing hypotheses, false findings, patch causality, reproducibility, novelty checks and independent reviewer evidence.
Explore the interactive capability atlas on cosmipher.com
Have a use case that crosses several areas?
Share the systems, evidence and workflow you want to evaluate. The same request form serves individual researchers and enterprise teams.
Request access for your use case →
Ways to put Freedom to work
Cosmipher's integration routes connect the model with security products, custom agent workflows and private endpoint arrangements.
| Route | Where it fits |
|---|---|
| Inside Cosmipher AI Security | AgentSPM, AI Firewall, AI Security Testing and AI Supply Chain workflows |
| Custom agent workflows | Domain-specific work connected to existing tools, tickets and approval paths |
| A private endpoint | Research, detection engineering, code review and incident analysis under access and audit controls |
| Enterprise engagement | A discussion of the organization's systems, security objectives and intended workflow |
Cosmipher works with IndiaMART and Angel One and participates in NVIDIA Inception, connecting its research with enterprise security and governance workflows.
Explore Cosmipher · Explore Cosmipher AI Security · Discuss access and integration →
Evidence-grounded security work
Evaluate security work through four assessment modes, from technical reasoning to source-grounded analysis and checked execution.
| Mode | Inputs and assessment focus |
|---|---|
| CHAT | A visible problem statement, technical reasoning and calibrated conclusions |
| ARTIFACT | Supplied source code, logs, configuration or other evidence |
| LAB | An authorized isolated environment with reproducible checks |
| AGENT | A model-plus-tool workflow with explicit permissions, execution traces and a verifier |
The six-dimensional security rubric evaluates technical correctness, evidence grounding, root cause, preconditions and impact, defensive usefulness, and uncertainty calibration. It connects a finding with the evidence, mechanism, repair and validation needed to evaluate it.
Benchmark performance
Freedom V1.1 ranks first in eight of 12 benchmark comparison cohorts, spanning cybersecurity, code repair, scientific reasoning and tool-assisted expert tasks.
Explore every result below, including complete six-model comparisons for all 12 benchmarks.
| Benchmark | Freedom V1.1 score | Position in comparison cohort | Task |
|---|---|---|---|
| CyberGym | 95% | 1 / 6 | Can it reproduce a vulnerability with working code? |
| SEC-Bench Pro | 82% | 2 / 6 | Can it find vulnerabilities in JavaScript engines? |
| ExploitGym | 40% | 2 / 6 | Can it build an exploit in a controlled environment? |
| DeepSWE v1.1 | 84% | 1 / 6 | Can it fix issues in real software repositories? |
| GPQA Diamond | 94% | 3 / 6 | Can it solve graduate-level science questions? |
| HLE with tools | 64% | 2 / 6 | Can it answer expert questions using tools? |
| CTIBench MCQ | 95% | 1 / 6 | Can it answer threat intelligence questions? |
| CTIBench RCM | 98% | 1 / 6 | Can it map a CVE to the underlying weakness? |
| CyberMetric 10K | 99% | 1 / 6 | How broad is its knowledge across nine cyber domains? |
| SecEval | 99% | 1 / 6 | Can it answer questions across nine security categories? |
| MMLU Computer Security | 99% | 1 / 6 | How well does it know core computer security concepts? |
| AthenaBench | 95% | 1 / 6 | Can it reason across six threat intelligence tasks? |
Complete peer comparisons
Open a benchmark to explore all six models and scores.
CyberGym — Freedom 95%
Can it reproduce a vulnerability with working code?
| Model | Score |
|---|---|
| Freedom V1.1 | 95% |
| XekRung 1.5 27B Preview | 88.92% |
| Gemini 3.8 Flash Cyber | 86.26% |
| ASL-Cyber-Flash | 85.87% |
| GPT-5.5-Cyber | 85.6% |
| GLM-5.3 | 84.5% |
SEC-Bench Pro — Freedom 82%
Can it find vulnerabilities in JavaScript engines?
| Model | Score |
|---|---|
| GPT-6 Astra | 85.4% |
| Freedom V1.1 | 82% |
| GPT-5.6 Sol | 79.1% |
| GPT-6.1 Sol | 78.8% |
| DeepSeek V4.1 Flash | 62.8% |
| DeepSeek V4 Pro | 56.4% |
ExploitGym — Freedom 40%
Can it build an exploit in a controlled environment?
| Model | Score |
|---|---|
| GPT-6 Astra | 42.4% |
| Freedom V1.1 | 40% |
| GPT-6.1 Sol | 35.1% |
| Claude Fable 5.1 | 30.4% |
| GPT-5.6 Sol | 30.3% |
| Claude Fable 5 | 28.4% |
DeepSWE v1.1 — Freedom 84%
Can it fix issues in real software repositories?
| Model | Score |
|---|---|
| Freedom V1.1 | 84% |
| Gemini 4 Argon | 77.9% |
| GPT-6.1 Sol | 75.2% |
| Claude Opus 5.5 | 74.2% |
| DeepSeek V4.1 Flash | 74.2% |
| GPT-6 Astra | 74.1% |
GPQA Diamond — Freedom 94%
Can it solve graduate-level science questions?
| Model | Score |
|---|---|
| GPT-6 Astra | 96% |
| Gemini 3.8 Flash | 95.3% |
| Freedom V1.1 | 94% |
| Claude Fable 5.1 | 93.7% |
| Kimi K3 | 92.9% |
| DeepSeek V4.1 Flash | 90.9% |
HLE with tools — Freedom 64%
Can it answer expert questions using tools?
| Model | Score |
|---|---|
| Claude Fable 5.1 | 65% |
| Freedom V1.1 | 64% |
| DeepSeek V4.1 Flash | 63.9% |
| Claude Opus 5 | 63.6% |
| Kimi K3 | 59.8% |
| GPT-6 Astra | 57.2% |
Freedom uses public-set retrieval in this HLE evaluation.
CTIBench MCQ — Freedom 95%
Can it answer threat intelligence questions?
| Model | Score |
|---|---|
| Freedom V1.1 | 95% |
| Claude 4.6 CyberXpert | 84.1% |
| GPT-5.4 | 78.5% |
| Gemma 4 31B | 74.4% |
| Qwen3.6 35B | 73% |
| GPT-OSS 20B | 69.3% |
CTIBench RCM — Freedom 98%
Can it map a CVE to the underlying weakness?
| Model | Score |
|---|---|
| Freedom V1.1 | 98% |
| RedSage 8B DPO | 75.7% |
| Claude 4.6 CyberXpert | 75.4% |
| GPT-5.4 | 74.2% |
| Qwen3.6 35B | 71.8% |
| Gemma 4 31B | 70.5% |
CyberMetric 10K — Freedom 99%
How broad is its knowledge across nine cyber domains?
| Model | Score |
|---|---|
| Freedom V1.1 | 99% |
| Claude 4.6 CyberXpert | 96.2% |
| GPT-5.4 | 96% |
| Qwen3.6 35B | 95.6% |
| Gemma 4 31B | 95.2% |
| Primus Nemo 70B | 93.4% |
SecEval — Freedom 99%
Can it answer questions across nine security categories?
| Model | Score |
|---|---|
| Freedom V1.1 | 99% |
| GPT-5.4 | 82% |
| Gemma 4 31B | 78.2% |
| Qwen3.6 35B | 74.2% |
| RedSage 8B DPO | 74.2% |
| Primus Nemo 70B | 72.8% |
MMLU Computer Security — Freedom 99%
How well does it know core computer security concepts?
| Model | Score |
|---|---|
| Freedom V1.1 | 99% |
| Claude 4.6 CyberXpert | 90% |
| Gemma 4 31B | 90% |
| Qwen3.6 35B | 88% |
| GPT-5.4 | 87% |
| GPT-OSS 20B | 86% |
AthenaBench — Freedom 95%
Can it reason across six threat intelligence tasks?
| Model | Score |
|---|---|
| Freedom V1.1 | 95% |
| GPT-5 | 66.1% |
| Gemini 2.5 Pro | 63.6% |
| GPT-4o | 58% |
| Gemini 2.5 Flash | 54% |
| GPT-4 | 51.4% |
Security Capability Atlas results
| Capability-atlas evaluation | Score | Scoring basis |
|---|---|---|
| 20-case Security Showcase | 97.00% | 582 / 600 rubric points |
| 431-scenario catalog | 92.34% | 11,940 / 12,930 rubric points |
The 431-scenario catalog records 378 pass, 51 partial and two fail dispositions. The 20-case showcase records 19 exact and one partial key match. Each fully scored item has six rubric dimensions, with scores from zero to five per dimension.
Long-context performance
| Context length | RULER score |
|---|---|
| 4K | 89.58% |
| 32K | 95% |
| 128K | 95% |
| 256K | 94% |
| 512K | 92% |
| 1M | 90% |
Freedom V1.1 supports a configured context ceiling of 1,048,576 positions, with RULER results spanning 4K through 1M context lengths.
Evaluation methodology
Freedom's scores come from Cosmipher's internal evaluations. Peer scores come from the sources cited in Cosmipher's benchmark comparison. Scoring follows each benchmark's task protocol, and tool-enabled results include the evaluation harness and environment. The Security Capability Atlas uses a six-dimensional rubric scored from zero to five per dimension.
Explore the complete benchmark results
Architecture and model specifications
| Property | V1.1 specification |
|---|---|
| Model | Cosmipher-Freedom-V1.1 |
| Developer | Cosmipher Tech Labs, India |
| Total logical parameters | 602B |
| Sparse active parameters | 8B during prefill and 16B during decoding |
| Modality | Image and text input, autoregressive text output |
| Architecture | Multimodal causal encoder-decoder mixture of experts |
| Backbone description | 20-layer causal encoder and 20-layer decoder |
| Sparse routing description | 384 routed experts, six selected per token, plus a shared expert |
| Configured context ceiling | 1,048,576 positions |
| Supporting subsystems | Engram memory, CSA2 indexing, mHC residual mixing, vision processing and multi-token prediction |
Cosmipher's development programme combines 10B-30B security-specialist models, cross-scale representation correspondence, transformations across tensor families and learned capacity expansion.
Cybersecurity knowledge and data design
Cosmipher's ~400 TB compressed cybersecurity knowledge collection supports specialist training, retrieval and evaluation across all five Vs of big data.
| Dimension | Data design |
|---|---|
| Volume | ~400 TB of compressed cybersecurity knowledge |
| Variety | CVEs, code, attack write-ups, papers, web content, video and transcripts |
| Velocity | Refresh cycles around new CVEs, exploit techniques and defensive guidance |
| Veracity | Source tracking, deduplication and separation of evaluation cases from training material |
| Value | Specialist training, evidence retrieval during investigations and realistic security testing |
The collection is organized into five source families with provenance and quality checks.
| Source family | Coverage | Provenance and quality checks |
|---|---|---|
| Technical repositories | CVE/CWE records, code, configurations, packet and log samples, detection rules | License, origin, deduplication, patch and version mapping |
| Cyber threat intelligence | Advisories, attack patterns, malware and fraud reporting | Time validity, attribution caveats and source reliability |
| Expert knowledge | Security engineering, defense and adversary tradecraft | Authorship, recency, technical fidelity and rights |
| Practitioner knowledge | Tools, labs and investigations captured as text | Provenance, language, accuracy and rights |
| Adjacent engineering | Operating systems, cloud, networks, compilers, cryptography and telecommunications | Relevance, technical correctness and recency |
Six security-training scenario families
Security training on supporting models transfers specialist expertise into Freedom. Training, calibration and held-out evaluation use separate example sets.
| Family | Training focus |
|---|---|
| Intent interpretation | Authorized requests, ambiguous intent and assistance requiring restriction |
| Secure-code review | Vulnerability identification, defensive fixes and technically correct explanations |
| Threat modeling | Assets, trust boundaries, attack surfaces and mitigations |
| Incident and artifact analysis | Suspicious artifacts and evidence triage inside approved labs |
| Access-control reasoning | Permissions, least privilege and policy-bounded actions |
| Counterexamples and abstention | Uncertainty, misleading evidence and situations requiring review |
Training examples shape model parameters. Calibration examples tune response decisions. Held-out cases measure transfer and generalization. Deduplication, near-duplicate screening and scenario-family separation keep these sets distinct.
Explore the data and training programme
Serving environment and integration
Freedom V1.1 runs as a private tensor-parallel service on four NVIDIA H200 SXM GPUs, with 564 GB aggregate HBM3e, served using vLLM 0.32. The software stack includes PyTorch 2.13, Python 3.12 and Ubuntu 24.04. Use the access form to define your deployment and integration requirements.
| Infrastructure | Specification |
|---|---|
| GPUs | 4 × NVIDIA H200 SXM |
| GPU memory | 564 GB HBM3e, 4 × 141 GB |
| Aggregate memory bandwidth | 19.2 TB/s |
| Parallelism | Tensor parallel across all four GPUs |
| System memory | 3 TB |
| Storage | 500 TB |
| Serving engine | vLLM 0.32 |
| Framework | PyTorch 2.13 |
| CUDA profile | Driver 12.8 · toolkit 13 |
| Operating system | Ubuntu 24.04 |
| Runtime | Python 3.12 |
Connect through an OpenAI-compatible API. Configure your endpoint and credential through environment variables.
import os
from openai import OpenAI
client = OpenAI(
base_url=os.environ["FREEDOM_BASE_URL"],
api_key=os.environ["FREEDOM_API_KEY"],
)
response = client.chat.completions.create(
model="Cosmipher-Freedom-V1.1",
messages=[{"role": "user", "content": "Hello."}],
max_tokens=256,
temperature=1.0,
top_p=0.95,
)
print(response.choices[0].message.content)
The same request is available through cURL.
curl "$FREEDOM_BASE_URL/chat/completions" \
-H "Authorization: Bearer $FREEDOM_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "Cosmipher-Freedom-V1.1",
"messages": [{"role": "user", "content": "Hello."}],
"max_tokens": 256,
"temperature": 1,
"top_p": 0.95
}'
Work with Cosmipher to configure context, modality and operational requirements for your endpoint.
Explore the model and infrastructure
Research citation and V2.0 roadmap
Use this citation in your research.
@misc{cosmipher_freedom_v11,
author = {Cosmipher},
title = {Cosmipher-Freedom-V1.1},
year = {2026},
note = {Security-focused multimodal Mixture-of-Experts model.}
}
The Freedom V2.0 roadmap targets 2T total parameters, with a 400 TB corpus available for preparation. Access the current Freedom V1.1 offering through the contact form.
Full security-domain catalog
AI and agent security — 9 domains and 74 domain scenarios
| Domain | Scenarios | Representative atlas tasks |
|---|---|---|
| AI agent security, agency and tool permissions | 10 | Agent goal hijack · Tool misuse |
| Prompt injection and jailbreak resilience | 10 | Retrieved document override · HTML comment injection |
| MCP tools, connectors and agent integrations | 8 | Tool schema trust · OAuth scope minimization |
| Multi-agent autonomy and delegated authority | 7 | Delegation chain audit · Agent role confusion |
| RAG, vector stores and knowledge-base safety | 8 | Cross-tenant vector leakage · Document ingestion provenance |
| AI model security and trustworthiness | 8 | Sensitive data disclosure · Instruction hierarchy |
| AI model, dataset and artifact supply chain | 8 | Model weight integrity · Adapter provenance |
| AI security testing, red teaming and evaluation validity | 8 | Benchmark contamination · LLM judge disagreement |
| AI inference and deployment security | 7 | Inference endpoint auth · Prompt/response logging privacy |
Application security — 9 domains and 72 domain scenarios
| Domain | Scenarios | Representative atlas tasks |
|---|---|---|
| Web authentication and session security | 8 | Session fixation analysis · Reset token validation |
| Web injection and interpreter boundaries | 8 | SQL injection code review · Stored XSS review |
| Web business logic and access control | 8 | IDOR / BOLA audit · Checkout price integrity |
| API security and service integrations | 8 | BOLA request set · BFLA admin path |
| Android and iOS application security | 8 | Android exported component review · iOS ATS configuration |
| Secure software lifecycle and code quality | 8 | Secure PR review · Threat modeling in design review |
| Memory safety and native code | 8 | Stack boundary check · Use-after-free lifecycle |
| Software supply-chain security | 8 | XZ-style build provenance · Dependency confusion review |
| Smart contract and DeFi security | 8 | Solidity access control · Reentrancy invariant |
Cloud and infrastructure — 11 domains and 82 domain scenarios
| Domain | Scenarios | Representative atlas tasks |
|---|---|---|
| AWS cloud security | 8 | IAM permission graph · Cross-account trust audit |
| Azure / Entra identity and security | 7 | Entra conditional access coverage · Azure RBAC scope review |
| GCP cloud security | 7 | GCP IAM inheritance · Service account impersonation |
| Kubernetes security | 8 | RBAC escalation review · Admission control design |
| Containers and virtualization | 7 | Dockerfile vulnerability review · OCI artifact provenance |
| Infrastructure as code and policy as code | 7 | Terraform privilege review · Ansible credential hygiene |
| Linux and Unix hardening | 8 | Sudoers review · Systemd persistence triage |
| Windows and Active Directory defense | 8 | Kerberos anomalous ticket triage · Privilege graph audit |
| Network protocols and segmentation | 8 | Packet timeline · Unexpected east-west traffic |
| Wireless, RF and telecommunications security | 7 | Spectrum whitelist classification · Wi-Fi enterprise configuration |
| Endpoint and browser security | 7 | Browser extension permissions · EDR event correlation |
Detection and response — 9 domains and 69 domain scenarios
| Domain | Scenarios | Representative atlas tasks |
|---|---|---|
| SOC triage and detection engineering | 8 | Sigma detection creation · Splunk SPL translation |
| Threat hunting and behavior analytics | 7 | Hypothesis-driven hunt · Beaconing versus scheduled jobs |
| Incident response and crisis management | 8 | Ransomware tabletop · Compromised cloud key |
| Digital forensics and evidence integrity | 8 | Disk evidence chain of custody · Memory snapshot triage |
| Malware analysis and detection engineering | 8 | Synthetic malware behavior report · Sandbox detonation summary |
| Reverse engineering and binary analysis | 7 | Disassembly control flow · ELF import analysis |
| Vulnerability analysis and prioritization | 8 | CVE exposure triage · Patch prioritization |
| Passive recon and external exposure | 8 | Certificate transparency interpretation · Passive DNS change analysis |
| Authorized penetration testing and purple teaming | 7 | Lab scope design · Proof-of-fix exercise |
Data, identity and cryptography — 6 domains and 42 domain scenarios
| Domain | Scenarios | Representative atlas tasks |
|---|---|---|
| Data security, privacy and governance | 7 | PII data-flow mapping · Database row-level security |
| Applied cryptography and secure protocols | 8 | AEAD nonce reuse · JWT/JWS key confusion |
| PKI, certificates and zero trust | 7 | Certificate authority compromise tabletop · Mutual TLS failure |
| Post-quantum and future cryptography | 6 | Crypto agility inventory · Hybrid key establishment |
| Fraud, abuse and identity verification | 7 | Account takeover anomaly · Transaction graph fraud |
| Email, phishing defense and collaboration | 7 | Phishing header triage · Business email compromise |
Industry, people and research — 7 domains and 52 domain scenarios
| Domain | Scenarios | Representative atlas tasks |
|---|---|---|
| Industrial control and OT security | 7 | PLC network zoning · Modbus telemetry anomaly |
| IoT, embedded and firmware security | 7 | Firmware update trust · Default credential inventory |
| Physical security and cyber-physical interfaces | 7 | Badge-access and identity correlation · CCTV data protection |
| Governance, compliance and third-party risk | 8 | NIST CSF control mapping · ISO 27001 gap assessment |
| Adversarial literacy and human factors | 7 | Phishing awareness classifier · Security training design |
| Security research methods and originality | 8 | Historical CVE versus novel bug · Reproducibility of findings |
| Security foundations and first-principles | 8 | Trust boundary inventory · Attack surface taxonomy |
Request access for your research or organization
Bring Cosmipher the security work you want to explore, whether that is an agent-permission boundary, an application-review workflow, a SOC investigation, cloud controls or a research programme spanning several domains.
All model-access requests, including corporate and individual requests, must begin with the contact form.
- Open the Cosmipher contact form.
- Answer the six questions about you and your team.
- Submit your request to start the access discussion.
Complete the access-request form →
Explore the company and its research · Explore Freedom V1.1
Sources and visual credits
Model specifications, capability coverage, benchmark data, training and deployment details are documented by Cosmipher. The links below provide the model reference, company information and access form.
Cosmipher · Freedom V1.1 · Access form
Technical reference — Cosmipher-Freedom-V1.1: Architecture, Development Methodology, and Security-Oriented Capabilities of a 602B-Parameter Mixture-of-Experts Model, Adarsh Keshri, Cosmipher Tech Labs, Technical Report v6.0, 11 November 2026.
The cover is an AI-generated illustration. Technical figures visualize Cosmipher's architecture, capability atlas and evaluation data.