AI security agent that finds vulnerabilities in AI-built apps, prepares fixes, and verifies the live app after deployment
Find and prevent secrets and prompt injection attacks in AI coding agent sessions
Stage
Startup
Not given
Founded
2026
Not given
Based in
US
Not given
Pricing model
Free trial
Open source
Pricing
Free scan with no card required. 7 days free trial with card, then $80/month USD. One organization with up to three included full-access people. $20/month per additional full-access person.
Not given
Key features
Continuous attack-surface mapping of exposed endpoints, services, and certificates
Vulnerability investigation with confirmed evidence and real impact explanation
Automatic fix pull requests on separate GitHub branches
Live verification that vulnerabilities are actually closed after deployment
Team notifications via Slack, Discord, and email with weekly summaries
Detection of exposed endpoints, leaked keys, open buckets, and injection paths
Domain ownership verification via DNS TXT record
Live retest to confirm vulnerability closure after fix deployment
Plain-English vulnerability titles with technical details in records
Multi-domain support on single surface map
Scan recorded AI agent sessions for known credentials
Guard against prompt injection attacks
Local console with audit trails at 127.0.0.1
Support for multiple AI coding agents including Claude Code, Cursor, and Codex CLI
Default policy with 21 rules across 19 action classes
Test policy against replayed attack scenarios
No model or network in the guard's hot path
What makes it different
Verifies vulnerabilities are actually fixed by testing the live deployed app after fixes ship
Prepares actionable fix pull requests you review and control before deployment
Designed specifically for AI-built applications that ship vulnerable code by default
Continuously monitors your changing app surface and checks again automatically
Shows evidence behind every finding so you can verify the work yourself
Deterministic policy-based enforcement rather than model-based decisions
Runs locally with no telemetry and no account required
Open source and free forever under Apache-2.0
Replayed 36 real attack scenarios with zero bypasses
Zoltra
Stroq