Trained on 40B+ tokens of CVE data, MITRE ATT&CK, threat intel, SIEM documentation, and incident response playbooks. Ships with 14 built-in security tools — enrich, correlate, contain, and document incidents autonomously, evidence preserved, human-in-the-loop on destructive actions.
Expert routing sends each query to the right slice of the MoE — threat intelligence, vulnerability analysis, incident response, compliance reasoning, and more — all within 2.6B active parameters.
Analyze logs, network flows, and behavioral signals. Maps every finding to MITRE ATT&CK tactics and techniques.
Enrich, correlate, contain, and document — end-to-end. Human approval gates on destructive actions.
CVE details in context. CVSS vectors, exploit-in-wild status, compatible patches, and prioritized remediation.
OWASP Top 10 and CWE-mapped findings with fix-ready diffs. Works on Python, JS/TS, Go, Java, Rust, C/C++.
Splunk SPL, Elastic KQL, Sentinel KQL, and Lucene — native query syntax, not pseudocode.
SOC 2, ISO 27001:2022, NIST CSF 2.0, PCI-DSS v4.0, CIS Controls v8, HIPAA. Audit-ready gap analyses.
Analyze packet captures, firewall rules, network anomalies, and DNS telemetry. NetFlow/IPFIX correlation.
Timeline reconstruction, artifact analysis, memory forensics guidance, and chain-of-custody documentation.
Test plans, payloads, and post-engagement reporting — only in authorized scope, refuses without scope declaration.
Every CyberAgent deployment ships with a pluggable tool layer. Destructive tools are gated behind policy-based human approval.
32 transformer layers, GQA attention (16 heads, 4 KV), and an 8-expert MoE with top-2 routing. Experts specialize during training across threat intel, vulnerability analysis, IR playbooks, code review, compliance, and tool orchestration.
Run on your own GPUs or air-gapped — threat-intel lookups route through on-premise feeds.
Production inference. Tensor-parallel, Hermes tool-call parser.
vllm serve \ thinkingdbx/cyberagent-moe-9b \ --dtype bfloat16 \ --max-model-len 32768 \ --tool-call-parser hermes
Single-command local run for dev machines and edge.
ollama pull \ thinkingdbx/cyberagent-moe-9b ollama run \ thinkingdbx/cyberagent-moe-9b
Transformers + bf16. Full tool schema via chat template.
AutoModelForCausalLM
.from_pretrained(
"thinkingdbx/…",
torch_dtype=bf16,
device_map="auto")
High-security environments — local threat-intel DB, no egress.
docker run -p 8000:8000 \ -e SIEM_ENDPOINT=… \ -e TI_DATABASE=/data/ti.db \ thinkingdbx/cyberagent