Security and Data Privacy, Answered Up Front
For the engineer or security reviewer who has to sign off. What runs where, what leaves your network, what an agent can reach, and what we haven't done yet. The answers match the ones we give in real security reviews.
Last updated October 2026. Where this page and your agreement differ, your agreement governs.
- Runs
- On your server, your cloud account, or fully offline
- Tenancy
- One company per copy. No shared database.
- Your data
- Never sent to us. Never used to train models.
- Models
- Your provider and your key, or a model on your own hardware
- Telemetry
- None in normal running. Updates only when you ask.
- Agents
- Act as one person. Every action checked.
- Record
- Every action logged, in your environment
- Source
- KickSide source-available; Wippy runtime MPL-2.0
- Patches
- Critical and high fixes within 30 days
- Incidents
- You hear from us within 24 hours
- Certified
- Not yet. SOC 2 in preparation.
Exactly What Leaves Your Network, and Why
KickSide is software you install. Nothing crosses your boundary unless you configured it to, so your firewall can block everything else.
- The application. One copy for one company, on your host.
- Your records. Contacts, deals, documents, knowledge and history.
- The database. SQLite to start, Postgres to grow, on your storage.
- The audit record and logs. Readable by your people without asking us.
- Model calls. To the provider you chose, on your own key. With a model on your own hardware, nothing leaves at all.
- Connections you add. HubSpot, Google, Slack and the rest, using credentials you issue and can revoke.
- Module installs and updates. From the Hub catalog, only when an administrator starts one.
- Your records, to us. We hold no copy and no standing access.
- Your data, to other customers. No shared environment exists.
- Your data, into training. Not ours, and not the model providers', whose commercial API terms exclude inputs.
- Telemetry. The copy doesn't phone home in normal running.
Need it air-gapped? Run a model on your own hardware, install the modules you need up front, and the copy runs with no outbound connection at all.
The Attacks a Reviewer Asks About, and What Stops Each One
AI-built software adds new ways for things to go wrong. These are the ones security teams raise most, and the control that bounds each.
| Threat | The risk | What stops it in KickSide |
|---|---|---|
| Prompt injection | The riskAn instruction hidden in an email, web page or document tells the agent to do something else. | What stops it in KickSideThe agent holds only the permissions of the person it acts for, and anything consequential waits in the approval inbox. An injected instruction can't reach past either. |
| Leakage through the model | The riskYour data ends up with a model vendor, or keys end up in a prompt. | What stops it in KickSideCalls go only to the provider you pick, on your key, or to a model on your own hardware. Connection credentials are never sent to the model. |
| Privilege escalation | The riskAn agent or automation reaches data its user couldn't. | What stops it in KickSideEvery action is checked on the server, against who is acting and what they touch. Hiding a button is never the control. |
| AI-written code | The riskGenerated code reads files, opens the network or touches other systems. | What stops it in KickSideIt runs in an isolated sandbox that reaches only the tools it was given. It starts with no files, internet or other parts of the system. |
| A bad module | The riskA package brings in code or schema changes nobody reviewed. | What stops it in KickSideEach install shows its publisher, license, version and database changes first, applies as one recorded change, and rolls back if any step fails. Anything from outside the catalog is marked unverified. |
| An AI tool overreaching | The riskClaude Code or Cursor connects with more reach than the person using it. | What stops it in KickSideEach connection carries one person's identity, is limited to named parts of the copy and chosen tools, expires, and can be revoked at any time. |
| The vendor | The riskWe're breached, or we disappear. | What stops it in KickSideWe hold no copy of your data and no standing access to your copy. It keeps running if we vanish, because nothing needs to check in with us. |
The Controls, in Detail
What's built into the platform, so every module and agent inherits it rather than reinventing it.
Identity and Access
- Named accounts with roles your administrators assign
- Permissions checked on every action, on the server
- Single sign-on through Google, Microsoft or any OIDC provider, so multi-factor sign-in follows your identity provider's policy
- Disable a user or revoke a connection at any time
Agents and AI
- Every agent and automation acts as a specific person
- A new agent starts with no access to anything
- Consequential actions wait for a person in the approval inbox
- Your provider and key, or Ollama, LM Studio or vLLM on your own hardware
- Revoke the model key and all AI processing stops at once
Isolation and Code Execution
- Isolated, supervised workers inside one program
- Sandboxed code, including R and Python, reaches only the tools it was given
- A failed piece restarts on its own, and long jobs resume where they stopped
Data and Network
- One company per copy, with no shared database
- Data sits on your storage, so disk and backup encryption follow your standard
- Run it behind your reverse proxy and terminate TLS there, like any internal service
- Outbound traffic only to what you configured
Audit and Logging
- Every run records what ran, as whom, which records it read, which model and what it did
- Stored in your environment, readable and exportable without asking us
- Application logs stay on your host, where your log agent can ship them to your SIEM
Change and Supply Chain
- Every capability is a versioned module
- Installs are previewed, recorded and rolled back on failure
- Out-of-catalog modules are flagged as unverified
- The code is readable: KickSide is source-available, and the Wippy runtime is open source under MPL-2.0
The Practices Behind Each Release
A security review asks about the company as well as the software. Here is how the team that builds KickSide works.
Still being formalized: the written policy set behind these practices, as part of our SOC 2 preparation. We'll tell you in your review which documents exist today.
Where We Stand Today
A false yes on a security form costs more than an honest no. Here is what's done, what's partial and what isn't done yet.
| Item | Status | Detail |
|---|---|---|
| SOC 2 | StatusIn preparation | DetailAudit expected in 2027 at the earliest. |
| ISO 27001, HIPAA | StatusNot held | DetailIf either is a hard requirement today, we'd rather tell you now than after the fourth call. |
| Independent penetration test | StatusNot yet | DetailYou're welcome to test your own copy. We'll support the test and share how we fix what it finds. |
| Patch timeline | StatusCommitted | DetailCritical and high: 30 days from disclosure. |
| Incident notification | StatusCommitted | DetailWithin 24 hours of confirming an incident affecting your copy or data. We'll write it into your contract. |
| Signed modules | StatusNot shipped | DetailModules from any source install and are marked unverified. Treat them like any code from the internet. |
| Automated dependency scanning | StatusPartial | DetailThe runtime's build runs automated security checks. Scanning across every module is being added. |
| Data processing addendum | StatusDraft | DetailFor deployments where we host your data. Current draft on request. |
| No-training commitment in contract | StatusDraft | DetailThe practice is in place today. The contract wording is being finalized, and we'll send the draft. |
| Source code you can read | StatusAvailable | DetailKickSide on GitHub, source-available. The Wippy runtime is open source under MPL-2.0. |
What You Get When You Review Us
This page is the summary. In a review, you get the evidence behind it.
Your Questionnaire, Answered in Full
Any partial answer says exactly what is and isn't in place, with a date where one exists.
A Walkthrough With Our Engineering Lead
Architecture, data flow and trust boundaries, with your security team, at any stage.
The Paperwork
The data processing addendum draft, the no-training wording and enterprise terms, on request.
A Copy to Probe
Install it in two commands on a machine you control, and test it however your policy requires.
Quick Answers for Security Teams
No. In normal running the copy sends nothing to us. The only contact with us is when an administrator chooses to install or update a module from the Hub catalog.
Only what you configure: your model provider (none if the model is on your own hardware), the services you connect, and the Hub catalog when installing or updating modules. Everything else can be blocked.
Yes. Connect a self-hosted model through Ollama, LM Studio or vLLM, install the modules you need up front, and the copy runs with no outbound connection.
Disable the user. Revoke a person's AI tool connection and that tool loses access. Revoke the model key at your provider and all AI processing halts, with no action needed from us.
In your copy, on your server. Connection credentials are used by the platform on the agent's behalf and are never placed in a prompt or sent to the model.
Everything lives in the database and the application's data directory on your storage, so your existing backup policy covers it. A restore needs the release version, those two and your configuration.
Yes. Test your own copy however your policy requires. Tell us what you find and we'll ship the fix and tell you what changed.
Use the form below and tell us it's a security review or a vulnerability report. A person from our team replies, usually within one business day. Confirmed critical and high-severity issues are fixed within 30 days.
We tell you within 24 hours of confirming it: what was affected, what we know about the cause, and what we're doing. Because the copy runs in your environment, your team handles containment at the infrastructure level, and we bring the application knowledge, help analyze logs and ship the fix.
Bring Your Questionnaire
Send us your security questionnaire or your list of questions. We'll answer it in full, as things actually stand, and put our engineering lead in front of your security team.
This site is protected by reCAPTCHA.
Rather probe it first? Install a copy in two commands.