01
The agent goes rogue
It does something nobody asked for: a payment, an email, a deleted record.
Our answer
Every agent has written rules for every kind of action: allowed, ask first, or never. It cannot act outside them or change them.
The control layer for AI agents
Your AI agents work. You stay in charge.
AI agents now pay invoices, write to clients, file documents and start other agents. The Companion, DCT-A by DynaGuard, makes sure every one of them, and every agent they start, acts only within the rules you set. Every important step waits for a person to approve it with a face or fingerprint, and you can require that approvals happen only on your phone.
Tested with ChatGPT and Claude: type “open Companion” and the Companion box opens inside the chat. Designed for any AI app or agent that supports remote MCP connectors with sign-in, and for your own code through REST.
What you type
Pay invoice INV-1001 from Gulf Steel TradingSUPPLIER-1 to account KW82 NBOK 0000 0000 0000 08ACCOUNT-2, and email the receipt to Mona Al-RashedPERSON-2 at mona@cedar.exampleEMAIL-3.
Details highlighted. The AI company would see all of them.
Payments: an approved payment becomes a signed payment order for your bank or finance team. The Companion never moves money itself.
The agentic age has a trust problem
Agents can act. That is exactly what makes them frightening.
Agents no longer just answer questions. They hold real access, work in teams and start new agents of their own. These are the fears we hear from every bank, business and practice, and how the Companion answers each one.
02
The agent is tricked
A hidden instruction inside an email, web page or document takes over the agent (prompt injection).
Our answer
A tricked agent is still held to your rules. Payments, emails and filings wait for an employee, and anything outside a sealed instruction is refused.
03
Too much power
The agent holds a master key to systems and money, far beyond what its task needs.
Our answer
Agents never hold lasting power. Each one gets short-lived, signed authority, limited to its task, with daily limits it never even sees.
04
Agents start agents
In multi-agent systems, power spreads down a chain of agents nobody can see.
Our answer
A helper agent always gets the same power or less, never more. Helpers share one limit, and stopping the lead stops them all.
05
Your data leaves
Client names, bank accounts and patient details flow to the AI company with every request.
Our answer
The AI sees tokens such as SUPPLIER-1 instead of real details. Real details are used only at the moment an allowed action happens.
06
No proof, no blame
When something goes wrong, nobody can show who allowed what, or what really happened.
Our answer
Every rule, approval and action is signed with post-quantum cryptography and recorded. Download a signed proof for your auditors.
And if you ever feel uneasy: one tap on Stop my agents ends every agent's authority at once and cancels anything still waiting. An action already being carried out at that moment finishes and is recorded in Activity.
Up and running in ten minutes
Create your organisation
Open the Companion on your phone or computer with your invitation code. Choose personal, business or bank, your country and currency, and create your passkey.
Connect your AI
Tap Connect an AI, choose your app and copy one address, or scan the code to open the steps on your computer. Pair your own agents with a short code: no secrets to copy. Choose what each one may do.
Approve what matters
Ask in plain words. Look-ups happen at once; anything important waits for your approval, then DCT-A carries it out and keeps the proof. Payments become a signed payment order for your bank or finance team.
Multi-agent systems
Agent teams, without a chain of hidden power
Lead agents can start helper agents to split up the work. With DCT-A, authority can only shrink as it moves down the team, never grow.
- Same power or less. A helper's rules are always the same as its lead's or narrower. Anything wider is refused.
- One shared limit. Helpers spend from the lead's daily limit, so a team cannot multiply its budget.
- You choose the size. Decide whether an agent may start helpers at all, and how many.
- One stop for the whole team. Stopping the lead stops every helper with it.
- Maker-checker for banks. Whoever connects an agent can never approve its requests.
Watch
See Companion in action
New · Start here
Then: let the AI work, you decide
Then, in three minutes
Start here · 60 seconds
Then, in three minutes
One feature at a time: watch the nine feature films.
Learn and try
Every feature, step by step. Then try it yourself.
Twenty-six short chapters, each a narrated film and an interactive walkthrough, recorded on the live demo. Or open the live demo: your own private organisation with example data and a demonstration AI. Nothing real is paid or sent.
Nine feature films
Your AI can act. You decide how far.
Each film shows one feature on the real product screens, in about a minute and a half. All records shown are synthetic demonstration data. Payments are signed payment orders for your bank or finance team to execute.
For every organisation and profession
Connect your own systems, such as case files, patient records, claims or customer records, and decide how each tool may be used.
Banks and financial firms
Maker-checker built in: whoever connects an agent can never approve its requests.
Businesses
Accounts payable and customer communication, with a second employee for every new supplier.
Insurers
Claims look-ups at once; settlements only after an adjuster approves.
Law firms
Case files and client letters; every court filing waits for a lawyer.
Clinics and hospitals
Patient details protected from the AI company, with your organisation's own key.
Individuals
Bills, paperwork and emails handled by your assistant, approved from your phone.
Extreme security for your text and files
Use any AI. Keep every name.
Names, identity numbers, accounts and addresses become tokens before anything reaches an AI, in text and in whole files, in English and Arabic. Add your own words too, such as a project code name, and protect whole paragraphs: highlight a confidential clause in Word, or choose it in the Companion. Each choice is checked first, and you are told if hiding it would spoil the AI's answer. The AI still does the work, and you read its answer or its file with the real details on your own device only.
Patent pending
Built so that nobody acts alone
Not the AI, not a stolen laptop, and not a single insider.
Post-quantum protection
Rules, authorities and approvals are signed with post-quantum cryptography, built on the newest international standards and designed to stay secure against future quantum computers.
Short-lived authority
Agents never hold lasting power. Their authority is brief, follows only the rules you signed, and ends the moment you stop them.
Sealed limits
Agents never learn your limits, approved suppliers or recipients, so they cannot probe for gaps.
Extreme security for your text and files
Names, ID numbers, bank accounts and addresses become tokens before anything reaches the AI, in text and in Word, Excel, CSV and PDF files. Add your own words, and protect whole paragraphs you highlight or choose. Use any AI app, then read its answer or its file with the real details on your own device only. Watch the films
Your own key
Protect your details with a key only your organisation holds. While it is locked, our servers cannot read them.
Separation of duties
Four-eyes for bank details, two approvals for large payments and record changes, and, if you choose, approvals limited to your phone.
Tested before every release
Every release passes end-to-end tests in real browsers and an internal security review with attack tests before it ships. An external assessment is planned. Security and trust
What the Companion does not do: it never moves money. An approved payment becomes a signed payment order (an ISO 20022 bank file and a PDF) that your bank or finance team executes.
For developers
Any framework that can call MCP (Model Context Protocol) or REST: LangChain, CrewAI, AutoGen, the OpenAI or Claude agent tools, or plain code. Pair your agent with the code shown in the Companion (python3 dcta_pair.py CODE "Agent name"); after the person approves on their phone, the agent receives its credentials directly, and no secret is ever shown on a screen. A lead agent can start helper agents with the same or narrower rules.
If something goes wrong: 401 means the token expired or the agent was stopped, so get a new token. 400 INVALID_INPUT means the body was not JSON with Content-Type: application/json. NOT_ALLOWED is final for that request: the reason is shown to the owner in Activity.
# 1. Get a short-lived access token
TOKEN=$(curl -s https://connect.dynaguard.ai/oauth/token \
-d grant_type=client_credentials \
-d client_id="$CLIENT_ID" -d client_secret="$CLIENT_SECRET" \
| sed -E 's/.*"access_token": ?"([^"]+)".*/\1/')
# 2. Look up unpaid invoices
curl -s https://connect.dynaguard.ai/v1/tools/list_invoices \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{"status":"unpaid"}'
# {"ok":true,"code":"OK","rows":[{"invoice":"INV-1001",
# "supplier":"SUPPLIER-1","amount":"USD 75.00",...}]}
# 3. Ask to pay it (waits up to 40 seconds for an employee)
curl -s https://connect.dynaguard.ai/v1/tools/pay_invoice \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{"invoice":"INV-1001","amount":"75.00"}'
# approved in time: {"ok":true,"code":"OK","status":"done",
# "result":{"order":"PO-...","status":"payment order issued..."}}
# not yet: {"ok":false,"code":"APPROVAL_REQUIRED",
# "request_id":"q..."}
# 4. Later, check a waiting request
curl -s https://connect.dynaguard.ai/v1/tools/check_request \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{"request_id":"q..."}'For investors
Let the AI work. You decide.
Companion · DCT-A by DynaGuard sits between AI agents and the actions they take. Organisations set the rules, employees approve the important steps on their phones, anything not approved is stopped, and every decision leaves signed proof.
- Live since 15 October 2026, with pilot clients; the Companion box inside ChatGPT and Claude since November 2026
- Works with any agent: ChatGPT, Claude, Gemini, Copilot, Llama, Mistral and an organisation’s own
- Multiple United States provisional patent applications filed
Portable Document Format (PDF). To talk to us, write to md@dynaguardtechnologies.com or watch the film.
Join the pilot
We are inviting a small number of organisations to use the Companion with their own agents. You get an invitation code, the user manual and direct support from our team.
Addresses and contact
Websitedynaguard.ai
The Companion (phone and computer)companion.dynaguard.ai
Desktop side panelcompanion.dynaguard.ai/panel
Connector for AI apps and agents
https://connect.dynaguard.ai/mcpPilots and enquiriesmd@dynaguardtechnologies.com
Company
DynaGuard Technologies LLC
Austin, Texas, United States
Kuwait branch
Dyna Guard Solutions
Kuwait