LLM Security Platform
Know what your deployed model actually does when it's attacked.
Run a versioned, self-hosted catalog of 22 security probes against your own Ollama or OpenAI-compatible endpoints, mapped to the OWASP Top 10 for LLM Applications 2025 — and turn the results into tracked findings, audit trails, and reports.
Platform features
A versioned test bench for the model you actually deployed
OWASP LLM Top 10 Coverage
22 automated tests span all ten OWASP Top 10 for LLM Applications 2025 categories, from prompt injection and system-prompt leakage to insecure output handling and excessive agency.
Self-Hosted, No SaaS
Runs entirely on your own infrastructure — no SaaS component, no telemetry, no account to create. Prompts and responses never leave your network.
SSRF-Hardened Target Testing
Every outbound request to a model endpoint is validated against a port allowlist and DNS-resolved address checks, with a hard, non-overridable block on cloud metadata addresses.
Versioned Policies & Fingerprinted Findings
Policies are immutable, versioned selections of which tests run and at what threshold, so a run from months ago stays reproducible. Findings are deduplicated by fingerprint across runs.
Findings Lifecycle
Every finding moves through a real review lifecycle — OPEN, ACKNOWLEDGED, RESOLVED, or ACCEPTED_RISK — with one-click retesting, so decisions are recorded, not lost in chat history.
Audit-Ready Reports
Export a versioned JSON for your own tooling, or a server-rendered PDF with a score/threshold verdict, findings table, evidence appendix, and framework coverage mapping.
Evidence Handling & Retention
Configure evidence capture per policy — hash-only, redacted preview, or encrypted-at-rest with audited decryption — with retention windows enforced automatically.
RBAC with TOTP MFA
ADMIN and READ_ONLY roles, TOTP-based multi-factor authentication, and forced password rotation plus MFA enrollment for every generated credential.
Latest updates
From the SOCFortress blog
Jul 29 2026
Introducing the SOCFortress LLM Security Platform
A self-hosted platform that runs a versioned catalog of 22 security probes against your own Ollama or OpenAI-compatible endpoints, mapped to the OWASP Top 10 for LLM Applications 2025.
Read on Medium
Aug 18 2026
I Asked a Local LLM to Repeat Its Instructions. It Said No, Then Did It Anyway.
A first run against a local Ollama model finds system-prompt leakage and stereotype generation in minutes — and an honest look at where the platform's own heuristics get it wrong.
Read on Medium
Ready to find out what your deployed model actually does under attack?
Request access Frequently asked questions
Search keywords..
What model endpoints can the platform test?
Does my prompt and model data ever leave my network?
What does a test result actually prove?
How is the test catalog kept up to date?
Didn’t find the answer you were looking for?
Contact us, we’re here to help