Enforce Security
at Generation Time
ReskSafety sits between your applications and your models as an OpenAI-compatible firewall: RBAC capability bitmasks, token-level filtering, editable policies, multi-provider routing, and a full admin console. Deployed in one Docker command.
Catch the Intent.
Block the Tokens.
Prompt filters are jailbreakable and post-hoc moderation acts too late. ReskSafety enforces security below the prompt — where no injection can reach.
Injection Blocking
Banned phrases are hard-blocked at the logits level: physically ungeneratable, not merely filtered.
Jailbreak Resistance
Multi-token Aho-Corasick matching catches rephrasings and partial matches across token boundaries.
Tool Call Gating
Missing capability bit? The tool's trigger phrases are banned before the call can even start.
Continuous Audit
Every request, tool call, and blocked content is logged and auditable from the admin console.
Block AI-Specific Threats
Every request passing through the firewall is filtered, scored, and logged. The console surfaces enforcement in real time:
1. JWT decoded → user roles + capabilities_mask
2. Tool check → bit 0 missing → 403
3. Policies compiled → banned phrases / token biases
4. Provider routed → OpenAI · vLLM · Ollama · custom
5. resklogits filter → hard (-inf) or bias (-penalty)
6. Response scanned → post-filter + session recorded
One Integration,
Full Pipeline Coverage
Two decoupled layers: the application layer knows users, roles, and policies; the resklogits engine knows token IDs only. The Policy → tokenizer bridge connects them.
Real-Time Enforcement
Every tool call, phrase, and model output is inspected during generation — before it can cause harm.
Policy Hot-Reload
Edit a YAML policy, and new rules take effect within seconds — no restart, zero downtime.
Per-User RBAC
64-bit capability bitmasks computed from roles: tools, code gen, DB access, PII, email — enforced per user, not per deployment.
Multi-Provider
One firewall in front of OpenAI, vLLM, Ollama, or custom gateways — selected per request.
Client (App / Agent)
│
▼
──── ReskSafety Firewall ────
│ JWT Auth (Argon2 + httpOnly)
│ RBAC bitmask check (64-bit)
│ Policy compilation
│ Tool gating (capability bits)
│ Provider routing (X-Provider-Id)
│ resklogits (Aho-Corasick, GPU)
│ Post-filter + audit log
│
│ Admin Console (React + TS)
│ Users·Roles·Policies·D3 graph
─────────────────
│
▼
LLM Provider
Tested Where It Matters
Attack Coverage
DAN injections, role confusion, system-prompt leakage, long-context manipulation, multi-turn attacks, and code injection — blocked at the token level.
PINT Framework
Validated against the Prompt Injection Testing framework: basic injections, complex multi-step strategies, extended contexts, and adversarial patterns.
"Prompt filters are UX. The real boundary is at the token — that's where ReskSafety enforces." RESK Security — Design Principle
Take Back Control
Full admin console, policy editor, and firewall in action — no installation needed.