com-kotobalabsEmpower those who protect.
Open-weight models of Mithril — specialized LLM inference for verified white-hat researchers.
Mithril (formerly Kotoba Cloud) is operated by Kotoba Labs Inc. The handle com-kotobalabs is unchanged, so existing model IDs keep working.
— Defensive by intent. Models help auditors and white-hat researchers review code, validate vulnerabilities, and prepare fixes.
— Honest, reproducible builds. Every model documents what changed, the method, the base model, and the serving footprint.
— Hardware-real optimization. Weights are quantized to what the target GPU actually runs — e.g. NVFP4 for NVIDIA RTX PRO 6000.
mithril.fund is where these weights are put to work. Identity-verified researchers walk a fixed workflow — INVESTIGATE → VALIDATE → REMEDIATE — from reviewing authorized code to a responsible-disclosure fix.
“One person’s discovery becomes a community’s strength.”
— Access: Passkey + post-quantum (ML-DSA-65) auth, Stripe Identity eKYC, AML/CTF screening, per-engagement scope confirmation.
— Free research allowance: 1,000 requests/day, up to 32,768 output tokens per request. Beyond it: plans from $20/mo or prepaid AI credits from $10.
— Supply chain: signed releases (Ed25519 + ML-DSA-65), distributed over content-addressed IPFS/CID.
A free public catalog of security datasets, served from the edge. No account, no key. Every file is CORS-open and supports HTTP Range, so duckdb and pandas read it directly.
— NVD CVE 2.0 full history (1999–today) as yearly Parquet, with CVSS 4.0/3.1/3.0/2.0, CWE and CPE counts — daily.
— CISA KEV (known exploited vulnerabilities) and FIRST EPSS exploit-probability scores — daily.
— MITRE CWE (weekly) and MITRE ATT&CK Enterprise STIX bundle (weekly).
— Row counts, byte sizes and sha256 are published in index.json and per-dataset manifest.json.
Try it (duckdb):
SELECT id, cvss31_sev
FROM read_parquet(
'https://knowledge.mithril.fund/nvd/parquet/nvd_cve_*.parquet')
WHERE cvss31_score >= 9.0;The Knowledge Graph & Ontology database of Mithril (formerly the Kotobase graph plane). Tenant-isolated graphs of datoms over immutable IPLD snapshots; content identity is a CID.
— Query languages: Datalog (Datomic-shaped EDN), Cypher, SPARQL, Gremlin bytecode, and bounded read-only SQL.
— Ontology: every entity is tagged with an ontology type; the workbench inspects observed attributes and types.
— For agents: an MCP server at graph.mithril.fund/mcp, and an HTTP API at api.mithril.fund/v1/database. Writes need an explicit Biscuit capability, not just a session cookie.
— Start free, no credit card: create a workspace and open a database right away.
Build and deploy bounded ontology applications: author a typed Mithril Form, then run one finite agent loop — OWL inference → SPARQL selection → SHACL validation → compile → conformance test → content-addressed deploy.
👉 mithril.fund — the platform
📖 docs.mithril.fund — documentation
📚 knowledge.mithril.fund — open security data
🗂️ graph.mithril.fund — graph & ontology database
🧪 app.mithril.fund — app studio
📰 blog.mithril.fund — blog
🧩 kotoba-lang.org — the language
✉️ support.mithril.fund · support@mithril.fund — support
Use of Mithril and these models is restricted to authorized security research on in-scope targets. Operator: Kotoba Labs Inc.