Security at Quint
What Quint installs, what it collects, where that goes and how it is protected. Written for the person reviewing Quint before it goes on a fleet. It states limits as well as protections. It is not a complete list: ask us about anything it leaves out.
What runs where
Quint is software on each machine and a console in the cloud. A signed package installs a daemon, an operating-system sensor and a network filter, and adds its hooks to the configuration files of the coding agents it recognises. Together they record what AI agents do on that machine and score it there, with no cloud round-trip in the decision.
Scoring on the machine uses deterministic rules, with no language model in the decision. The console's Ask and policy-drafting features do use one: what someone types there, the earlier turns of the conversation and the records Ask reads go to Amazon Bedrock.
Quint's cloud runs in AWS us-east-1. The console's pages are served by Vercel, and its requests to Quint's API, the responses and the session cookie pass through Vercel's network. Each machine registers with your organization and sends what it records.
What Quint records and sends
By default the daemon sends what it captures to Quint's cloud: prompts and model responses, tool arguments and results, command lines, file paths and network destinations. That includes what people type into an agent, the contents of files an agent reads and the output of commands it runs.
Some of what it records is not an agent's. Requests from any program on the machine to the AI provider hosts Quint intercepts pass through its proxy and are captured the same way when it can decrypt them. When a browser opens an AI chat site, Quint uploads a record of the site and the process id, not the content. Its network filter sees the destination of every other outbound connection from every process, and uploads only those of recognised agent sessions. It also reads the MCP servers that agents' configuration files declare.
By default, credentials the detector recognises are masked on the machine before upload, on some capture paths, not all. Everything else is sent unmasked, personal data included.
Two settings keep that content out of Quint's storage. With log_bodies: false in the daemon policy, the daemon replaces prompts, tool arguments and results with a marker before any event ships. With content suppression, which Quint enables for an organization on request, Quint's ingest removes the same fields before they are stored, apart from times when ingest cannot read the setting or the setting has just changed. Either way the cloud still receives metadata: event type, tool, agent, timestamps, findings, the working directory, session name, file and program names, network destinations and inventory reports.
The audit database on each device keeps tool arguments and results, and the prompts and payloads that agent telemetry reports, whatever log_bodies says.
Where your data lives
- Region: captured data is stored in AWS us-east-1, in the United States. Ask requests may be processed by Amazon Bedrock in us-east-1, us-east-2 or us-west-2.
- Tenancy: one cloud shared by every customer organization, isolated per organization by Postgres row-level security. That separates customers from one another. It is not a separate encryption key for each organization, and it does not restrict Quint's own administrative access to the database.
- Encryption at rest: the production database is encrypted with a key held in AWS KMS, and object storage is encrypted on the server side.
- Encryption in transit: Quint's API accepts TLS 1.2 and TLS 1.3 only.
- Network: the production database has no public address.
- Retention: the docs list how long each class of data is kept, and name classes with no automatic deletion yet (not a complete list), under Workspace and retention.
The service providers Quint uses are listed in the Privacy Policy.
What Quint can do on a machine
To read an agent's traffic to its model provider, the daemon terminates TLS on the machine, with a certificate authority it generates there and adds to the machine's trust store. Quint uploads the certificate, not the private key. The key stays on the machine, where an administrator of that machine could read it. The certificate is name-constrained to a built-in list of AI provider and coding-assistant hosts and domains; one entry is all of amazonaws.com, because models hosted on AWS are served from there.
The proxy adds a cookie and headers of its own to provider responses on the machine, to identify the agent, and removes them from requests before they leave.
By default the endpoint package blocks nothing an agent does. Rules written in the console do not reach the device yet, and enforcement at the agent-hook layer is off unless an operator turns it on. One command does act on a machine: a console administrator can end a session, and the daemon, on its next check-in, stops that session's processes.
The audit record
On each machine, the daemon signs each audit record with that machine's Ed25519 key and chains it to the previous record with SHA-256, so editing a record's tool, arguments, result, verdict or risk makes it fail verification. You can check a machine yourself, without access to Quint's infrastructure.
Four limits matter in an audit. Other columns, such as the session id, are not signed. An administrator on that machine can re-sign records with the key stored there. Deleting the newest records is not detectable. And a signature shows that the daemon wrote a record, not that what it describes happened: any local process can send the daemon agent-hook events.
Quint's cloud keeps a separate hash chain over each event's type, tool name, decision, risk score and time. It is not signed, it does not cover the event's content, and you cannot verify it yourself today.
Signing in to the console
People sign in with a one-time code sent by email, with Google, or through your identity provider by single sign-on. Quint does not add a second factor of its own: the email code is a single factor, and with Google or single sign-on your identity provider's rules apply.
Certifications
Quint has not had a SOC 2 audit and has no SOC 2 report.
The console maps each finding to the controls it touches in frameworks such as SOC 2 and NIST AI RMF. That mapping describes your agents' activity. It does not certify Quint.
Reporting a vulnerability
Write to security@quintai.dev with what you found and how to reproduce it.
Questions
For a security review, a questionnaire or anything this page does not answer, write to hello@quintai.dev or book a demo.
The product documentation has a page for security reviewers and the network requirements: the hosts and ports the software on each machine connects to, and how it behaves behind a proxy or TLS inspection.