Try to get past obsvr

See how obsvr evaluates every AI agent action in real time, from interception to audit record.

Real deterministic policy engine, running in your browser.
AGENT ACTION PAYLOADeditable JSON
⌘/Ctrl + Enter to evaluate
OBSVR GATEWAY - DETERMINISTIC EVALUATION
Evaluating...
Paste your agent code, see what's governed and what's exposed.
YOUR AGENT CODE
EXPOSURE REPORT
Paste code on the left, or click "Load Example."
policy_decisions.log
RECENT POLICY DECISIONS
PASS RATE: 67%BLOCK RATE: 33%
All evaluations stream securely to your CISO Compliance Ledger.
EXPORTABLE
[14:31:02] ACTION: aws_s3_read | STATUS: PASSED
LATENCY: < 1ms
[14:30:58] ACTION: send_email | STATUS: BLOCKED | REASON:
LATENCY: < 1ms
[14:30:45] ACTION: filesystem.read | STATUS: PASSED
LATENCY: < 1ms
[14:30:31] ACTION: extract_json | STATUS: PASSED
LATENCY: < 1ms
[14:30:22] ACTION: update_crm | STATUS: BLOCKED | REASON:
LATENCY: < 1ms
[14:30:15] ACTION: fetch_pricing | STATUS: PASSED
LATENCY: < 1ms
[14:29:58] ACTION: shell_exec | STATUS: BLOCKED | REASON:
LATENCY: < 1ms
[14:29:44] ACTION: read_db | STATUS: PASSED
LATENCY: < 1ms
[14:29:30] ACTION: generate_report | STATUS: PASSED
LATENCY: < 1ms
[14:29:18] ACTION: delete_record | STATUS: BLOCKED | REASON:
LATENCY: < 1ms
[14:29:05] ACTION: web_scrape | STATUS: PASSED
LATENCY: < 1ms
[14:28:51] ACTION: translate_text | STATUS: PASSED
LATENCY: < 1ms
Agent Simulation

Secure Your Agent

Compare native agent execution against obsvr-governed execution to see policy enforcement in action.

1. SELECT THREAT SCENARIO
Native LangChain
VULNERABLE
AGENT INITIALIZATION
import OpenAI from 'openai'
const client = new OpenAI()
const res = await client.chat.completions.create({
model: 'gpt-4o',
messages: [{ role: 'user', content: userInput }],
tools: [sqlTool, searchTool],
})
// tool call executed directly - no validation
await executeTool(res.tool_calls[0])
TERMINAL OUTPUT$ ./RUN_AGENT.PY
Waiting for execution...
LangChain + obsvr
SECURED
AGENT INITIALIZATION
import OpenAI from 'openai'
import { obsvr } from '@obsvr/sdk'
// one-time setup
obsvr.init({ environment: 'production' })
const client = obsvr.wrap(new OpenAI(), { user_id: req.user.id })
const res = await client.chat.completions.create({
model: 'gpt-4o',
messages: [{ role: 'user', content: userInput }],
tools: [sqlTool, searchTool],
})
// tool call is intercepted, policy-checked, and audited first
await executeTool(res.tool_calls[0])
TERMINAL OUTPUT$ ./RUN_AGENT_SECURED.PY
Waiting for execution...

Default Governance Policies

Every AI agent action is evaluated through configurable policy controls. No LLM inference required. Pure code-based enforcement.

1

PII Protection

Detects sensitive data in prompts, responses, and tool outputs. Block, redact, hash, or review before release.

2

Tool Governance

Restrict which tools agents can invoke. Prevent shell access, database writes, or external API calls outside policy.

3

Human Approval

Require manual approval before high-risk actions like payments, emails, production changes, or customer communications.

4

Model Routing

Force sensitive requests to approved models only. Block unsupported providers or fallback behavior.

5

Prompt Injection Defense

Detect jailbreaks, prompt injection, and malicious instructions before they reach your models or tools.

6

Compliance Controls

Enforce your organization's policy rules with configurable, versioned governance controls.

7

Cost Governance

Apply per-user, per-agent, or per-tenant token budgets and spending limits automatically.

+

Custom Rules

Define your own policy rules with YAML or JSON. Any check, any condition, any action.

No login required. This is a client-side simulation of obsvr's policy engine. The real engine adds about 45 microseconds for full governance.

12

Deploy a control plane for your LLM and agent stack in under five minutes.

Control what your LLMs and agents are allowed to do. See obsvr intercept, enforce, audit, and export evidence from real LLM and agent traffic.

Live event capture from your LLM calls and agent runs
PII detection running on real prompt examples
A fully-generated SOC 2 evidence packet (ZIP archive)
Your specific control gaps and how to close them

We'll respond within one business day to schedule a time that works. Prefer email? hello@obsvr.dev