Active Evals 12,482
Latency0.07ms
Last 60s Blocks47
Pass Rate68%

AI Models Guess. Exogram Enforces.

The 0.07ms deterministic firewall for autonomous agents.

Exogram does not fix outputs. It makes failures deterministic so your systems can.

Agent Action

"Extract pricing tiers into strict JSON."

LLM Output
{
  "tiers": [
    { "name": "Starter", "price": 29 },
    { "name": "Pro", "price": 99 },
    { "name": "Enterprise",  },
  ]
}

Returns malformed JSON with a trailing comma and missing required field.

Exogram Output
HTTP 403 ForbiddenBLOCKED:Schema mismatch(0.07ms)
ERROR BUBBLED:

ParseError: Missing required field "price_tier"

Client Layer
Client Behavior: Retrying with structured extraction prompt...
Final Status: Output validated and passed.
Execution Trace
1

Intercepted Tool Call

Payload Size: 412b

2

Python Logic Gates

Compute Time: 0.07ms

3

Bubbled Error to Client

Deterministic Response

4

Hash Committed

e3b0c442... → async ledger

Last evaluated payload size: 412 bytesEvaluation time: 0.07msResult: BLOCKED
Global Enforcement Ledger
Pass Rate: 68%Block Rate: 32%
🛡️ All evaluations stream securely to your CISO Compliance Ledger.Exportable
Agent Simulation

Build an NVIDIA NemoClaw Agent

Test how the Exogram API intercepts and validates generative payloads before the NemoClaw framework executes them on your infrastructure.

1. Select Threat Scenario

Native NemoClaw

Vulnerable
Agent Initialization
import nemoclaw

agent = nemoclaw.Agent(
  model="nemotron-4-340b"
)
Terminal Output$ ./run_agent.py
Waiting for execution...
> Initializing NemoClaw Agent [nemotron-4-340b]

NemoClaw + Exogram API

Secured
Agent Initialization
import nemoclaw
import exogram

agent = nemoclaw.Agent(
  model="nemotron-4-340b"
)
agent.use_authority(exogram.ExecutionGuard())
Terminal Output$ ./run_agent_secured.py
Waiting for execution...
> Initializing NemoClaw Agent [nemotron-4-340b]
Interactive Sandbox — Try It Yourself

Agent Simulation — Select a Payload

Exogram Gateway — Deterministic Evaluation

Select a payload or write your own to test the gateway

0.07ms
Compute Latency

70 microsecond median — production-verified across 5,000 payloads

8
Policy Rules

Deterministic logic gates — zero LLM inference

0
False Negatives

Zero false negatives in 5,000-payload red-team flood test

137 RPS
Throughput

200 concurrent connections, 0.00% failure rate

The 8 Deterministic Policy Rules

Every AI agent action is evaluated through 8 deterministic logic gates. No LLM inference. No probability. Pure Python code.

1Conflict Resolution

Blocks writes if unresolved contradictions exist in the semantic ledger. Ensures fact consistency before any state mutation.

2Quota Enforcement

Tier-based monthly evaluation limits. Free: 500/mo, Pro: 50K/mo, Developer: pay-as-you-go, Enterprise: unlimited.

3Transmission Guard

Agents draft messages — humans send them. Blocks all direct email.send, message.send, and sms.send operations.

4Billing Cap

Blocks billing operations exceeding $10,000 threshold. Prevents AI agents from authorizing high-value financial transactions.

5Compute Execution

Blocks destructive code: os.system, subprocess, rm -rf, eval(), exec(), __import__, shutil.rmtree, and more.

6Filesystem Guard

Blocks writes to system paths: /etc/, /root/, /var/, /usr/, /bin/, ~/.ssh. Prevents privilege escalation.

7API Exfiltration Gate

Blocks outbound API calls to untrusted domains. Only whitelisted domains (exogram.ai, supabase.co, googleapis.com) pass.

8Destructive DB Guard

Blocks DROP TABLE, TRUNCATE, DELETE FROM, ALTER TABLE. Prevents SQL injection and database destruction.

Works With Every Agent Framework

Exogram sits between your agent and execution. Use any framework — the governance layer is the same.

LangChain
LangGraph
CrewAI
AutoGen
OpenClaw
OpenAI Assistants
Anthropic Claude
Google Vertex AI
LlamaIndex
NeMo Guardrails
Hugging Face Agents
MCP (Model Context Protocol)
Custom REST Agents

Frequently Asked Questions

What is the Exogram Proving Ground?+

The Exogram Proving Ground is a free, interactive sandbox where developers fire adversarial payloads at Exogram's deterministic AI governance engine. It runs the exact same 8 policy rules as the production system — same code, same logic gates — but with zero database writes and zero authentication required. Completely stateless and safe for public use.

How fast is Exogram's policy evaluation?+

Exogram evaluates AI agent actions in 70 microseconds (0.07ms median). The 8-rule deterministic policy engine uses pure Python logic gates with zero LLM inference. In a 5,000-request production flood test under 200 concurrent connections, the engine sustained 137 RPS with a 0.00% failure rate. Every evaluation in the Proving Ground displays exact compute latency.

What is deterministic AI governance?+

Deterministic AI governance means every security decision is made by code — Python logic gates — not by AI models. Given the same input, you get the same output every time. No hallucination, no probabilistic uncertainty, no model inference in the decision path. This is fundamentally different from output filtering tools like Guardrails AI, which use LLMs to classify outputs.

How is Exogram different from Guardrails AI, Lakera, or NeMo Guardrails?+

Guardrails AI and similar tools validate model outputs after generation — they are output filters. Exogram governs before execution with deterministic logic gates, SHA-256 state hashing, and cryptographic execution tokens. Guardrails uses LLM-based classification (probabilistic). Exogram uses zero LLM inference (deterministic). Filtering outputs is not the same as governing actions.

Can I use Exogram with LangChain, CrewAI, AutoGen, or OpenClaw?+

Yes. Exogram integrates with any agent framework via REST API or MCP (Model Context Protocol). It sits between the agent and execution, intercepting every tool call for governance evaluation. Your framework decides what to do. Exogram decides what is permitted.

Do I need to sign up to use the Proving Ground?+

No. The Proving Ground is completely free — no login, no API key, no credit card, no signup required. It is a public demo endpoint at exogram.ai/proving-ground.

What attacks can I test?+

The Proving Ground includes 10 pre-built attack vectors: DROP TABLE SQL injection, rm -rf filesystem wipe, subprocess spawn, SSH key theft, data exfiltration, bulk email spam (5,000 recipients), unauthorized message sending, and more. You can also write custom JSON payloads with any action_type and payload.

Is this the real Exogram production engine?+

Yes — the Proving Ground runs the exact same 8 deterministic policy rules as production. Same Python logic gates, same blocked patterns, same SHA-256 state hashing. The only difference is that it skips database writes and audit logging, making it safe for unauthenticated public traffic.