Security at Kalmia
Customers send Kalmia the traces their production AI agents produce. Those traces can contain prompts, tool results, and whatever the agent saw, so protecting them is the core of the job. This page describes the controls we run today, not the ones we plan to. Last updated November 8, 2026.
How we approach security
Kalmia is a small team, so we lean on a few principles rather than a large apparatus. Every workspace's data is isolated at the database layer, not just in application code. Secrets are stored hashed or encrypted, never in plain text. Production access is limited to the people who operate the service, and the changes they make are logged. Dependencies are scanned continuously and patched against fixed deadlines. And we write down what we do, so that what this page says and what the system does stay the same.
Compliance program
We run our security program on Vanta, which continuously monitors our cloud accounts (AWS, Vercel, and Supabase), our code repositories, and our identity provider against the SOC 2 Security criteria and alerts us when a control drifts. We are preparing for a SOC 2 audit; when the report is available, customers can request it under NDA at team@kalmia.dev.
Our written policies are reviewed and approved by management at least annually. They cover information security and acceptable use, access control, asset management, business continuity and disaster recovery, cryptography, data management, human resource security, incident response, operations security, physical security, risk management, secure development, third-party management, and a code of conduct. Every employee accepts the policies that apply to them and completes security awareness training.
Infrastructure
Kalmia runs on Vercel (application hosting and serverless compute) and Supabase (Postgres database and authentication), with model inference on Amazon Web Services Bedrock and OpenAI. All of these providers hold their own SOC 2 or ISO 27001 attestations and encrypt data at rest. We operate from the United States, and data is stored and processed there. We do not run our own servers; there is no production infrastructure to reach other than through these providers' authenticated control planes.
Every response from kalmia.dev carries a baseline set of security headers: HTTP Strict Transport Security with preload, a Content Security Policy that forbids framing and plugin embedding, no-sniff and referrer controls, and a restrictive permissions policy. The application does not advertise its framework or version.
Data protection
- Encryption in transit and at rest. All traffic uses TLS. Our hosting and database providers encrypt data at rest.
- Tenant isolation. Every table that holds customer data is protected by Postgres row-level security keyed to the workspace. A query from one workspace cannot return another workspace's rows, regardless of application code.
- API keys. Workspace API keys are shown once and stored only as SHA-256 hashes. We cannot recover a key; if it is lost, it is rotated.
- Integration credentials. Tokens for Slack, Microsoft Teams, GitHub, PostHog, and Braintrust connections are encrypted before they are written to the database and removed when the integration is disconnected.
- Retention. Traces and everything derived from them are deleted 90 days after capture by an automated daily job. Deleting a workspace deletes its data immediately. Customers can ask us to delete specific traces sooner.
- No training on customer data. We do not train models on customer traces, and we use our model providers under business terms that forbid them from doing so.
Access control
Customers sign in with a one-time email link or a password; we never see passwords in plain text. Workspace owners control who is a member of a workspace and what role they hold, and can remove members at any time. Internally, production access is restricted to the engineers who operate the service and is granted on the principle of least privilege. Cloud root accounts are not used for day-to-day work, and any use of one is flagged by our monitoring and must be justified. Sensitive settings changes in the product are recorded in an audit log.
Secure development
- Code review. Since November 3, 2026, every change to the main branch requires a pull request approved by someone other than its author. Merges without approval are flagged by our compliance monitoring and must be justified.
- Automated checks. Every pull request runs linting, type checking, and the test suite, and deploys to an isolated preview environment for review.
- Dependency scanning. GitHub Dependabot scans every manifest in our repositories. We fix high-severity advisories within 30 days of detection, medium within 60, and low within 90. Our compliance monitoring tracks each advisory against these deadlines.
- Secrets. Secrets live in the hosting provider's encrypted environment configuration, never in the repository.
- Abuse controls. Per-workspace rate limits and daily spending caps bound what any single key can do, so a leaked key cannot run up unbounded analysis.
AI processing
Kalmia's analysis is performed by large language models from Anthropic (through Amazon Web Services Bedrock) and OpenAI. Trace contents are sent to these providers only to produce an analysis, under business API terms that prohibit training on the data and limit retention to short-term abuse monitoring. We run the analysis inside our own infrastructure; prompts and model outputs are stored with the workspace that owns the trace and deleted with it. The analysis is advisory: it is designed to surface failures for a human to verify, not to take action on its own.
Monitoring and incident response
Application errors are captured in Sentry and request logs in Axiom, with alerting on new and recurring failures. Our compliance monitoring watches cloud configuration, repository settings, and identity continuously. We maintain an incident response plan that defines roles, severity levels, and communication steps. If we learn of a breach affecting customer data, we notify the affected customers without undue delay and in any case within the timeframes our agreements and the law require.
Availability and continuity
The application is stateless and runs on our hosting provider's redundant infrastructure, so a single failed instance does not take the service down. Our database provider takes daily backups. Our business continuity and disaster recovery plan sets recovery objectives and the steps to restore service, and names who is responsible for each.
Subprocessors
The providers that process customer data on our behalf, and what each one does:
- Vercel: application hosting and serverless compute.
- Supabase: database and authentication.
- Amazon Web Services (Bedrock), Anthropic, and OpenAI: model inference for trace analysis.
- Upstash: rate-limit counters, keyed by workspace id.
- Resend: transactional email.
- PostHog: product analytics and interface session recordings, with input fields masked.
- Sentry and Axiom: error reports and request logs.
Services a customer chooses to connect (Slack, Microsoft Teams, GitHub, PostHog, Braintrust) receive data only as that customer configures. We will give customers notice before adding a subprocessor that handles customer data. The full list and the data each one sees are in our Privacy Policy.
Reporting a vulnerability
If you believe you have found a security issue in Kalmia, email team@kalmia.dev with enough detail to reproduce it. We will acknowledge your report within two business days, keep you informed as we work on it, and credit you if you wish once it is fixed. We ask that you give us a reasonable time to fix the issue before disclosing it, avoid accessing other customers' data, and not run automated scanning or denial-of-service tests against the production service. We will not pursue legal action against researchers who follow these guidelines in good faith. Our machine-readable disclosure policy is at /.well-known/security.txt.
Contact
Security questions, questionnaires, and report requests: team@kalmia.dev. How we handle personal information is described in the Privacy Policy, and the agreement that governs the service is in the Terms of Service.