AI agent security at Arclyx.
AI agent security in Arclyx starts with access: an agent uses only the tools you grant, through your own sign-in, and emails only you. Every run is recorded step by step, code runs in an isolated sandbox with no internet, and models are reached through providers that never train on your data, with zero data retention where it's offered.
- Read 14 unread emails
- Scanned the email text
- Tried a link from an email
- Counted senders in code
- Drafted 3 replies
- Emailed the summary
What are the security risks of AI agents?
An AI agent acts on its own with real access, so the risks are what it can reach, what it reads, what it spends and where its output goes. Too much access, prompt injection and runaway runs top most lists. Here is each risk and what Arclyx does about it.
- Only the tools you grant, through each person's own sign-in
- Outside content is scanned and fenced off as untrusted data
- Email only to the person it runs for; one Slack channel; drafts in Gmail and Outlook
- Plan limits checked before every run; a ceiling on every run
- An isolated sandbox with no internet and no credentials
- Every run recorded; a run that stops short waits for you
How does Arclyx limit what an agent can access?
By what it is given, not by what it is told. Each limit below is enforced by Arclyx outside the model, so an agent can't talk its way past one, even when a web page or an email tries to steer it.
Only the tools you grant
An agent is given only the tools you grant it, and you can remove any from its draft. Ready-made jobs mostly read and draft rather than send.
Your own sign-in
Built-in tools connect through each person's own sign-in to that service, so an agent works with that person's access, not more.
Delivery it can't change
Email goes only to the person the run is for. Slack posts go to the channel set on the agent; the model can't pick another.
Code with no way out
Code an agent writes runs in an isolated sandbox: a read-only system, no credentials and no internet access.
Only links it found
A page is opened only if the run's own search found it or you gave the link, and never at a private network address.
Tools that change are held
A connector's tools are checked when they're added. A tool whose description later changes, or hides instructions, is quarantined.
Where is my data stored, and is it used for training?
Arclyx is hosted in North America, and we don't use your content to train AI models. Models are reached through OpenRouter, with zero data retention where the provider offers it. Others, such as Anthropic and Google, keep prompts and responses for a limited period, usually around 30 days, only to detect misuse. None of them train on it.
- North America. Arc10 Technologies is a Canadian company.
- Through OpenRouter, with zero data retention where the provider offers it
- We don't use your content to train AI models
- Each organisation's data is kept apart; document search starts inside yours
- Encrypted over HTTPS
- Delete conversations and documents in the app; a deleted organisation's data goes after 30 days
How are connector sign-ins and API keys protected?
Connector sign-ins are encrypted before they're stored, with a key kept outside the database, and only Arclyx's servers use them to call the tool. API keys are kept only as a hash. The sandbox where agent code runs gets no credentials at all.
- Sign-in tokens encrypted before they're stored; you can disconnect at any time
- Stored only as a hash, never as the key itself
- Each has its own secret, stored encrypted and checked on every call
- Held by our sign-in provider, Supabase, not by us
- Entered directly with Stripe; card details never reach our servers
- Gets no credentials, and can't reach the internet to use one
Who can see which documents?
Access tags decide it. Admins tag people and documents, and a member's search, and their agents' search, only reaches documents whose tags match their own. What someone asks and what their runs say stays theirs: admins see usage and run status, not the content.
- Members
- Search documents whose access tags match their own, and so do their agents.
- Admins
- Set each person's and each document's tags, and see the organisation's shared documents.
- Personal documents
- Seen only by the person who uploaded them.
- Runs and chats
- A run's content is seen only by the person who ran it. Admins see each member's usage and each run's status, not what was said.
How do you check what an agent did?
Open the run. Arclyx records every run step by step: what the agent decided, each tool it called and what came back, the tokens and the cost. A run that needs you doesn't fail quietly; it waits on Home.
Every step
Each round, each tool call and its result, in order.
Tokens and cost
For the run and for each model call.
Guard events
When a check fires, such as fenced instructions, the run records it.
Needs you
A run that stops short shows on Home. A schedule that fails five times in a row pauses itself and tells you.
How do I report a security issue?
Email hello@arclyx.ai with what you found and how to reproduce it. A person reads every message. Please give us time to fix an issue before you share it. If a breach affects your personal information, we'll tell you and regulators as the law requires.
What I found:
The page or feature, the steps to reproduce it, and what you expected.
Questions about AI agent security.
Does Arclyx train AI models on my data?
No. We don't use your content to train AI models, and neither do the model providers we use. Where a provider offers zero data retention, we use it; others keep prompts and responses for a limited period, only to detect misuse, and longer if content is flagged.
Where is Arclyx hosted?
In North America. Arc10 Technologies is a Canadian company; some of our service providers may process data in other countries, as the privacy policy explains.
Can an agent send email as me?
It drafts, and you send. In Gmail and Outlook an agent writes email as drafts for you to review and send. Its own emails, from Arclyx, go only to the person it runs for.
Can an agent spend without limit?
No. Your plan's limits are checked before every run starts, every run stops at its ceiling ($2 unless you change it), and admins can set a monthly limit for each member.
What happens if an agent reads a malicious web page or email?
Content from the web, email and other outside sources is scanned for known injection patterns and fenced off as data, links are opened only if its own search found them, and even a steered agent can only email you and post where you set.
Where to go next.
Create an agent that stays inside its access.
One sentence in, a working agent out. Join the beta and we'll set you up.