Your agents never hold your API keys. NeoGriffin encrypts, proxies, and scans every call — 150+ detection patterns, autonomous threat analysis, full audit trail.
Proxy mode or SDK — your choice. Same protection, your architecture.
POST proxy.neogriffin.dev/proxy/chat { "model": "gpt-4o", "messages": [{ "role": "user", "content": "..." }] } → Scanned, proxied, audited. Zero key exposure.
Every week, another incident proves the same pattern: an AI agent with direct access to API keys, credentials, or production systems — and no security layer in between.
An employee connected a third-party AI tool with full OAuth access to their company workspace. The tool was breached — attacker escalated to the workspace and exposed production environment variables.
In 96% of agentic misalignment tests, the model attempted to leverage access beyond its intended scope. Model-level safety alone is insufficient.
AI coding agent deleted an entire production database in 9 seconds. The developer wasn't at fault — the agent simply had credentials it shouldn't have held.
Published research documented an 86% attack success rate against AI agents via prompt injection. NeoGriffin was built to close exactly that gap.
NeoGriffin is a transparent proxy. External agents connect to our endpoint — we store their API keys encrypted (AES-256-GCM), scan every request, and forward only clean calls to the LLM provider.
Libraries run inside the agent's process — if the agent is compromised, so is the library. NeoGriffin sits between the agent and the LLM provider. The agent has no access to credentials, no direct API connection, and no way to bypass the security layer.
Every input and output passes through 150+ detection patterns covering:
→ Prompt injection (jailbreaks, role hijacking, context manipulation)
→ Credential exfiltration (API keys, BIP-39 seed phrases, tokens)
→ Social engineering (authority impersonation, urgency exploitation)
→ On-chain threats (EVM & Solana blacklists, wallet monitoring, contract analysis)
→ Velocity attacks (anti-crescendo, rate-based threat escalation)
Every feature exists because a real attack or vulnerability demanded it.
150+ patterns covering jailbreaks, role hijacking, context engineering, and novel attack vectors. Tested against real-world payloads.
API keys stored AES-256-GCM encrypted. Agents never see or hold credentials. Zero key exposure architecture.
AI-powered pipeline that classifies threats, provides feedback, and improves detection accuracy autonomously.
700+ address blacklist across EVM and Solana with calldata decoding. Protects on-chain agents from interacting with known malicious contracts and wallets.
SHA-256 hash-chained logs for every scan, decision, and transaction. Tamper-evident by design. Thousands of entries and counting.
Anti-crescendo detection stops escalating attack patterns before they succeed. Rate-based threat escalation with automatic banning.
Use the proxy for vault mode, or integrate the Python SDK for local scanning without changing your base_url. Works with any provider, including Ollama.
Not benchmarks we wrote ourselves. External tools, external audits, real adversarial testing.
Maximum threat detection rating on external security evaluation.
Zero detections across 60+ antivirus engines. Clean infrastructure, no false flags.
Full score on adversarial penetration testing across all attack categories.
Across 150+ patterns with autonomous subagent feedback loop for continuous improvement.
From 29 open security items to zero. Every vulnerability identified and resolved.
Published research showed 86% prompt injection success against unprotected agents. NeoGriffin was built to close that gap.
NeoGriffin is live. We're onboarding a limited number of teams with dedicated support, direct feedback channel, and priority features.
Request Access on X →Contact @dagomint on X · Limited spots available