Get Started
Security and trust
Where your robot data goes, who and what can act on it, and how your team checks each finding the agent makes before acting on it.
This page answers two questions that a security reviewer and an engineering lead ask before they adopt Alloy. Can we trust Alloy with our robot data? Can we trust what the agent tells us?
Each claim below has a scope. If a claim applies only to some organisations, the page says so.
Summary
| Claim | Scope |
|---|---|
| SOC 2 Type II audited. | Alloy |
| HIPAA supported. A Business Associate Agreement (BAA) is in place before any US protected health information (PHI) is uploaded. | Organisations that handle PHI |
| Your organisation is the boundary. Your cases and data are kept separate from other organisations. People and agents see only the data of an organisation they belong to. | Every organisation |
| Agents act as the signed-in person. An AI tool that connects over MCP signs in with that person's Alloy account. | Every MCP connection |
| The agent's own replies on a Jira ticket post directly. Other ticket changes are made only after a person agrees. | Organisations that connect Jira |
| Findings link to their evidence. For MCAP recordings, a link can open Replay at the moment. | Reports and chat on the Agent Platform |
| Reports keep a version history. You can see what changed and restore an earlier version. | Every Agent Platform organisation |
| Dedicated environments and data-residency options are available. | By arrangement |
Certifications
- SOC 2 Type II: audited. Ask Alloy for the report as part of the security pack.
- HIPAA: supported. Alloy can support HIPAA requirements. If your data includes US protected health information (PHI), ask Alloy for a Business Associate Agreement (BAA). Sign it before you upload that data.
Where your data goes
1. Data comes in
Your robot data reaches Alloy in one of these ways. You choose which ones to use.
| Route | What happens |
|---|---|
| Browser upload | A person uploads recordings in the Alloy app. |
| Alloy Edge (Mesh Storage) | A small program on the robot or base station uploads each finished recording. It needs outbound HTTPS only, and it can remove sensitive data on the device before upload. |
| Python SDK (Mesh Storage) | Your scripts upload recordings with an API key that a person creates in the app. |
| Your cloud bucket | Alloy reads new recordings from an S3 or Google Cloud Storage bucket. You create the access role, and Alloy sets up the connection with you. |
2. Alloy stores and processes it in your organisation
Alloy turns each upload into a mission that the agent can search, query and report on. Missions, reports, scenarios, comments and team memory belong to your organisation. Other organisations cannot see them.
3. The agent reads the tools you connect
When you connect a tool, the agent can read from it to answer a question. You choose which tools to connect.
| Tool | What the agent uses it for |
|---|---|
| Slack | Questions and support threads in channels where you add the agent, and direct messages people send it |
| Jira | Tickets in the projects you connect |
| Confluence | Your triage guides and runbooks |
| GitHub | Your code at any branch, tag or release |
| Grafana, ClickHouse | Metrics, logs and telemetry around a failure |
A person on your team connects Jira, Confluence, GitHub, Grafana and ClickHouse in the organisation settings (organisation name → Manage → Integrations). Alloy switches on Slack with you during onboarding. Each integration page has a Permissions and privacy section; start at Integrations.
4. The agent uses AI models
The agent uses AI models to read your data and write its answers. Alloy chooses and tests these models. See Model testing.
5. Results go where you send them
Results stay in the Alloy app unless you send them out. They can also go to:
- the Slack channels you choose for recurring reports, and your own email address for scheduled-report emails;
- Jira, as the agent's replies on tickets, and other ticket changes after a person agrees (see What can change your systems);
- your own AI tools, through MCP, under the signed-in person's access.
Where your data lives
Your organisation runs in Alloy Cloud, or in a dedicated environment by arrangement. If your data must stay in a particular country or on your own infrastructure, we work with you on a setup that fits. The security pack covers where data is stored and processed.
Who and what can act on your data
People
Each person signs in to an organisation. They see the data of that organisation only.
AI agents (MCP)
Your team can connect its own AI tools to Alloy over MCP. Examples are Claude, Codex and Cursor.
- The agent signs in as a person. A person approves the connection in the browser. The agent then has that person's access, and no more.
- Alloy checks membership on every call. When a person leaves the organisation, their agents lose access.
- Your products and integrations decide the tools. An organisation sees only the tools for the products and integrations it has.
- Some actions stay with people. An agent cannot create SDK or SQL API keys, connect integrations, change billing, or pause or delete recurring reports. No MCP tool deletes recordings, missions or reports. On Mesh Storage, an agent can approve devices and create a development device key to test Alloy Edge.
- Hand-off to the agent. In some organisations a connected agent can hand a task to the agent in Alloy. See what the hand-off can do.
The agent
The agent runs inside your organisation with the access that your organisation gave it. In Slack, it works only in channels you add it to and in direct messages people send it. Add it only to channels whose members should see your organisation's robot data.
What can change your systems
| Action | Control |
|---|---|
| Reply in a Slack support channel | Alloy and your team set each role to respond to every new post (it can stay silent when nothing is needed) or only when @mentioned. Either way, every conversation in a channel with a role is recorded as a case. |
| Reply on a Jira ticket | The agent's own replies post directly. On Jira Service Management, they are internal notes. |
| Create, edit fields on, move or comment again on a Jira ticket | The agent proposes the change in Alloy or Slack, and makes it only after a person agrees. Ticket create, edit and transition: Alloy switches this on with you. |
| Create or update a report or recurring report | Allowed for people and for connected agents, under the signed-in person's access. |
| Approve or reject a device, change Alloy Edge configuration, create a development device key | Allowed for people and for connected agents on Mesh Storage, under the signed-in person's access. Set these tools to ask first in your AI tool. |
| Hand a task to the agent from a connected agent | Runs under the signed-in person's access. The agent can write reports, rename missions and edit their descriptions, create, edit or archive scenarios, save Replay layouts, and add or remove its own comments. Set the hand-off tool to ask first. See the hand-off. |
| Delete recordings, missions, reports, keys or devices | People only, in the app. |
How you can check the agent's work
Trust in an AI answer comes from evidence that you can check. Alloy shows its evidence so your team can check it quickly.
- Open the evidence. Findings link to their source: the moment in the recording, the chart, or the data behind it. For MCAP recordings, a link can open Replay at that moment with the relevant panels.
- Read the report history. Each report keeps a version history. See what changed, and restore an earlier version.
- Check each scenario match. Each match shows the reason it matched, its time window and the mission it came from.
- Agree before Jira changes. Beyond its own replies, the agent proposes ticket changes and waits for a person.
- Discuss it in place. Comment on a report or a mission. Rate answers with thumbs up or down.
- Ask for the basis. Ask the agent "what evidence supports this?" or "what did you check and rule out?". Ask it to separate what the data shows from what it infers.
Check the evidence before you act on a finding that affects safety or a release. The agent shortens the path to the evidence. It does not replace your engineer's decision. See Where Alloy stops.
In our security pack
Ask Alloy for the security pack before your security review. It covers certifications, where data is stored and processed, encryption, retention, export, model training and sub-processors.