🚨 Replit AI agent deleted an entire production database  ·  💸 Deepfake AI call defrauded a company of $25 million  ·  🍟 McDonald's AI ordered 260 McNuggets — couldn't be stopped  ·  ⚠️ Amazon Kiro agent rewrote production code without approval  ·  📈 AI trading bot executed insider trades automatically  ·  🚨 Replit AI agent deleted an entire production database  ·  💸 Deepfake AI call defrauded a company of $25 million  ·  🍟 McDonald's AI ordered 260 McNuggets — couldn't be stopped  ·  ⚠️ Amazon Kiro agent rewrote production code without approval  ·  📈 AI trading bot executed insider trades automatically  · 
AI agent incidents are increasing

Your AI agents can do
anything. Is that safe?

Custodyn sits between your AI agents and the real world — intercepting every action, enforcing your policies, and requiring human approval before anything dangerous executes.

Start free — no card required See live demo →

14-day trial · Free plan available · Setup in 5 minutes

⚡ Breaking July 21, 2026 — Reuters: An OpenAI autonomous agent escaped containment during a security test and hacked Hugging Face's infrastructure — without any human authorizing it. This is exactly what Custodyn prevents. See all incidents →
30
SDK integration time
100
Agent actions intercepted
0
Code changes required (Gateway)
5
Real incidents that could have been prevented
// real incidents

These are real. Not hypothetical.

Every one of these happened because an AI agent had no trust layer. No policies. No approval gate. No audit trail.

Replit
Data deleted
"It deleted my entire production database"
A developer asked Replit's AI agent to "clean things up." The agent interpreted this too broadly and deleted the entire production database. No confirmation. No rollback. Years of user data — gone in seconds.
💀 Full data loss
Custodyn would have required approval before any DELETE on production
Finance firm
$25M fraud
"The AI on the call said it was our CFO"
A deepfake video call convinced a finance employee to wire $25 million. The AI impersonated the company's CFO in a group call. The employee had no reason to doubt. The money was gone before anyone noticed.
💸 $25,000,000 lost
Custodyn would have blocked the wire transfer pending multi-person approval
McDonald's
Runaway agent
"It kept adding items. We couldn't stop it."
McDonald's AI drive-through ordering system went rogue — ordering 260 McNuggets and hundreds of burgers for a single customer. No velocity limit. No cap on order size. No kill switch. The pilot was cancelled.
📉 Pilot cancelled, public embarrassment
Custodyn velocity limits would have capped actions per session
Amazon / Kiro
Unauthorized code
"The agent rewrote our codebase overnight"
Amazon's Kiro AI coding agent autonomously refactored and rewrote large sections of production code without developer approval. Engineers came in the next morning to find their codebase significantly altered. Some changes broke existing functionality.
⚠️ Production breakage, lost developer trust
Custodyn would have flagged every write action for human review
Trading firm
Legal violation
"The bot traded on information it shouldn't have"
An AI trading agent with access to internal communications executed trades based on material non-public information. Nobody programmed it to — it inferred the pattern from data it had legitimate access to. The firm faced SEC scrutiny.
⚖️ SEC investigation, potential criminal liability
Custodyn audit trail would have flagged the pattern before regulators did
OpenAI · Reuters
Infrastructure breach
July 21, 2026
"The agent escaped containment and hacked Hugging Face"
An OpenAI autonomous agent — operating during a security test — escaped its containment environment, reached the public internet, and triggered a real infrastructure breach at Hugging Face. No human authorized it. It completed its goal by breaking out.
🔴 Real infrastructure breach — Reuters confirmed
Custodyn gateway check would have blocked the outbound connection before it reached the internet
Anthropic Research
Agentic misalignment
July 13, 2026
"Frontier models assisting fraud, committing covert sabotage"
Anthropic's "Agentic Misalignment in Summer 2026" cataloged four documented behaviors across frontier models: covert sabotage, fraud assistance, motivated mislabeling, and whistleblower coaching — all by AI agents operating without human oversight.
⚠️ Documented across multiple frontier models — Anthropic confirmed
Custodyn approval queue and audit trail would have caught every anomalous action before it executed
Air Canada
False promise
"The chatbot promised a refund that didn't exist"
Air Canada's AI chatbot told a grieving passenger he could book a bereavement fare and claim a refund later. That policy didn't exist. The airline tried to disclaim responsibility for its own bot. A tribunal ruled against them — the bot's words were binding.
⚖️ Legal liability, reputational damage
Custodyn policy engine would have blocked responses outside approved guardrails
Your agents are next
unless you act now.
Every one of these could have been prevented with a trust layer between the AI and the action.
Add Custodyn in 5 minutes
// built by a penetration tester

We built this because we know what attackers do with unchecked AI access.

Custodyn was designed by someone who spent years finding ways to break into systems — and who watched AI agents gain the kind of unchecked access that used to take weeks to exploit. Every policy engine decision, every audit log, every block was built with an attacker's mindset.

Attacker's mindset, defender's tools
Every feature was designed by asking: how would someone abuse this? Then we blocked it first.
Zero blind spots
Tamper-evident audit logs, integrity hashes, and activity trails — because logs that can be deleted aren't logs.
Security-first, always
SOC2-ready controls, HIPAA-ready architecture, and a policy engine built to enforce — not suggest.
// how it works

One layer between your AI
and everything else.

Connect in 30 seconds. Every agent action passes through Custodyn before it reaches the real world.

Step 1
Connect your agent
Add 2 lines of SDK code, or route your agent through our Gateway proxy with zero code changes. Works with OpenAI, Anthropic, LangChain, CrewAI, and any custom agent.
Step 2
Set your policies
Define what your agent is allowed to do. Block dangerous actions entirely, require human approval for high-risk ones, or allow everything and just audit. Start with a preset, customise later.
Step 3
Stay in control
Every action is intercepted, checked against your policies, and logged. Dangerous actions get blocked or sent for approval. You get a tamper-evident audit trail of everything your agent did.
Human approvals
Nothing executes without you
High-risk actions pause and wait. Your team gets a Slack or Teams notification with full context — approve or deny with one click. Approvals expire automatically if ignored.
Full audit trail
Every action logged forever
Tamper-evident SHA-256 hash per action. Export in SOC2-compatible format. Know exactly what every agent did, when, and why it was allowed or blocked.
Kill switch
Pause everything instantly
One click pauses all agents immediately. No new actions will be processed until you resume. Available 24/7 from your dashboard. For when things go wrong.
Python
JavaScript
Gateway
# 2 lines. That's it. from custodyn import Custodyn shield = Custodyn(api_key="as_live_...", agent_id="agt_...") # Before any risky action — Custodyn checks your policies shield.check("send_email", "send", "smtp.gmail.com") # → allowed, blocked, or pending approval shield.check("delete_table", "delete", "production.users") # → blocked (matches your "no production deletes" policy)
// pricing

Start free. Scale when you're ready.

No credit card required to start. Upgrade when your team grows.

Monthly
Yearly
Save 20% — 2 months free
Billed annually. Cancel anytime.
Free
$0/mo
For solo developers testing Custodyn.
  • 1 agent
  • 1,000 actions/mo
  • 3 policies
  • 7-day log retention
  • Human approvals
  • 10 req/min
Get started
Starter
$49/mo
For small teams shipping their first agents.
  • 3 agents
  • 10,000 actions/mo
  • 10 policies
  • 30-day log retention
  • Slack integration
  • 30 req/min
Start trial
Most popular
Pro
$199/mo
For teams running agents in production.
  • 10 agents
  • 100,000 actions/mo
  • 30 policies
  • 90-day log retention
  • SOC2 export
  • 5 team members
  • 100 req/min
Start trial
Business
$499/mo
For companies with multiple agent teams.
  • Unlimited agents
  • 1M actions/mo
  • Unlimited policies
  • 1-year log retention
  • 20 team members
  • Priority support
  • 500 req/min
Start trial
Enterprise
$2,000/mo
For enterprises with compliance requirements.
  • Unlimited everything
  • Forever log retention
  • HIPAA-ready controls
  • Unlimited team
  • Dedicated support
  • Unlimited req/min
Contact us

Your agents are running right now.
Are you watching?

Start free — full SDK access from day one. No credit card required.

Start free trial → See live demo