Security
Effective date: October 1, 2026
This document describes controls that are applied. It states what is implemented, not what the result of implementing it will be. Nothing on this page is an undertaking about an outcome. No method of transmitting or storing data over a public network is free of risk, and a document that said otherwise would be describing something other than software.
It is published by AIVIBE SOFTWARE SERVICES PRIVATE LIMITED and covers AicippY: the applications you install, the console at www.aicippy.com, and the services those reach. It is read with the Privacy Policy, which states what personal data is held and why, and with the Terms of Use, which govern where this document and the Terms address the same subject.
Where your code goes, and where it does not
AicippY is an agent that runs on your own machine. Your repository, your working tree and your shell are local to it. The product does not clone, mirror, index or upload your codebase, and there is no server-side copy of it.
What leaves the machine, and only this, is the following.
- What the agent sends to a model. The prompt, the conversation and whatever file contents the agent reads in order to answer are sent to the model provider serving that request. That is the working of the product, not an incident of it: an agent that cannot read the file cannot change the file. Which provider receives it is determined by the model you select.
- Account and billing records. Sign-in, entitlement and payment records, described in the Privacy Policy.
- Metering metadata. Where a request is served through the Company's own model proxy, the request is recorded as metadata — the model identifier, input and output token counts, whether it streamed, the outcome, the stop reason, the latency and a request identifier. Prompt text and model output are not columns in that record.
What does not leave the machine:
- Your repository as a repository. Nothing is cloned or synchronised to a server here.
- Files the agent did not read. The product has no background indexer and no crawler over your disk.
- Command output, diffs and session history, beyond what is sent to the model as part of answering.
- Tracing. Spans for model calls exist behind an experimental setting that is off unless you turn it on, and they are emitted to a collector you configure.
What the product does not do
Three positions are stated plainly, because a reader assessing this product is better served by them than by silence.
- The agent is not sandboxed. AicippY executes shell commands, reads and writes files and reaches the network with the privileges of the account it runs under. The permission system — the prompts asking you to confirm a command or a write — is a control over awareness and intent. It is not an isolation boundary and is not designed as one. Where true isolation is required, run AicippY inside a container or a virtual machine.
-
Server mode is opt-in and unauthenticated unless you set a password. Where it is
enabled, setting
AICIPPY_CODE_SERVER_PASSWORDapplies HTTP Basic authentication to it. Without that variable the server runs without authentication and warns that it is doing so. Restricting access to it is the operator's responsibility. - What you configure is outside this boundary. A model provider you point the product at with your own key, an MCP server you connect, and your own configuration files are yours. Their behaviour is governed by their own terms.
Credentials and tokens
- On your machine. The desktop application holds the signed-in session through the operating system's own credential encryption, in the main process only; the renderer asks for it across an inter-process boundary and never holds the file. The encrypted file is written with owner-only permissions. Where the operating system offers no credential store — a fresh Linux session with no keyring — the application says so rather than writing the value in the clear.
- Provider keys and command-line sessions. Keys and tokens written by the command-line client are stored in the application's own data directory with owner-only file permissions, applied on every write.
- Sign-in. Sign-in is through a managed identity directory. Tokens are RS256 JSON Web Tokens. Redirects use PKCE, with a single-use verifier and a state value that is checked on return and then discarded.
- In the browser. The console holds the sign-in token in session storage and discards it when the tab closes. No refresh token is held in the browser.
- Verification. Every token presented to a backend is verified against the directory's published signing keys. The key is selected by key identifier, the algorithm is required to be RS256, and the issuer, the token use, the application client and the expiry are each checked against fixed expectations. A clock skew of thirty seconds is allowed and no more. Signing keys are re-fetched at most once a minute for an unknown key identifier, so a flood of invalid tokens does not become a flood of key requests.
- Server-side secrets. Secrets are separated by purpose and held in a managed secret store, read at runtime. The key used to sign unsubscribe links is kept apart from the payment credentials.
Multi-tenant scoping
The platform database is shared with the other products of the group. Records are scoped on two axes, and the application's data access requires both.
-
Per-application scoping (
app_id). Every table holding this product's records carries anapp_idcolumn constrained by a foreign key to the application registry. The application resolves that identifier from the registry row at query time, as a scalar subquery inside the statement, rather than trusting a value carried in the environment — a value in the environment and a row in the database are two declarations of one fact, and on the day they disagreed the wrong one would be the one scoping a query about somebody's money. A start-up consistency check compares the two, and the registry row prevails. -
Per-tenant scoping (
tenant_id). Where a sign-in carries an organisation or tenant claim, that value is taken from the verified token and written and read with the record. Role assignments are keyed on the combination of tenant, product, user and role. - Row-level security in the database. The platform database is a multi-tenant PostgreSQL instance with row-level security enabled, and tenant isolation is enforced by policies held in the database itself rather than by the application that queries it. Policies are attached across the schema and restrict rows to the tenant established for the session, covering reads, inserts, updates and deletes. On the subscription records, row-level security is additionally forced, so the restriction applies to the table's owner as well and cannot be sidestepped by the connection that created it.
-
Application scoping on top of it. Reads and writes are also issued through
application SQL that carries the application and account scope on every statement, in the
WHEREclause rather than as a check performed afterwards. A query that cannot return another customer's row is the version of that rule which cannot be skipped by a later edit. This sits above the database policies, not instead of them: a defect in one is met by the other.
Encryption
- In transit. Requests to the published aicippy.com hosts are redirected to HTTPS, the edge negotiates TLS 1.2 as a minimum, the connection from the edge to the origin is TLS 1.2, and HTTP Strict Transport Security is sent with a two-year maximum age and subdomains included. It is not submitted to the browsers' preload list, because that submission is close to irreversible and is a separate deliberate act rather than something a deployment performs. Realtime connections use WebSocket over TLS in production. Connections from the application to the platform database are encrypted.
- At rest. Data is held in managed storage services — the managed relational database, object storage and key-value storage named in the subprocessor table of the Privacy Policy — and is encrypted at rest by those services under their own configurations. That is a statement about those providers, under their terms, and not a configuration held in this product's code.
Least privilege and network position
- The platform database is reachable only from inside the private network. It has no public endpoint. Administrative operations against it are invoked directly against the function that holds that permission, over an authenticated control-plane call, and are not reachable through any public URL.
- The inference role names the permitted foundation model resources directly. The model proxy accepts a closed list of model identifiers, by exact match, and no prefix, alias or case variation of them.
- The model proxy applies a rate limit of sixty requests of burst, refilled at one a second — above an agent loop running parallel subagents, and far below a script replaying a captured token.
- The proxy accepts only tokens issued by this product's own directory pool and application clients. The accepted client list is additive, so a migration between clients has a window in which both are valid and a customer's session is not the place that discovers otherwise.
- Production infrastructure is marked as retained, against removal by a deployment.
Integrity of payment messages
- Every notification from a payment processor is verified by HMAC-SHA256 over the exact bytes received, compared in constant time, before anything in it is acted on.
- The delivery is recorded with the raw payload, the signature presented, the verification result and the algorithm used, under a unique idempotency key that rejects a repeated delivery.
- The browser callback returning from checkout is verified by HMAC over the identifiers it carries.
- Redirects back from a payment flow are restricted to an allowlist of the processor's own domains.
- Card details are not held by the Company at any point.
The web surface
- A Content Security Policy is enforced at the edge and restricts which origins may supply scripts, styles, images, fonts, frames and network connections. The only third-party script origin permitted is the payment processor's checkout. It was first published in report-only form and moved to enforcement afterwards, because a policy one directive short does not degrade — it takes the payment form off the page.
- A Permissions Policy denies browser capabilities the Services do not use, and grants the payment capability only to the payment processor's own origins.
- Cross-Origin-Opener-Policy, Cross-Origin-Resource-Policy, X-Content-Type-Options, X-Frame-Options and a strict-origin-when-cross-origin referrer policy are set on responses from the edge.
- Server and infrastructure identification headers are removed from responses at the edge, and the account, billing and callback surfaces are served with a no-index directive.
Logging and audit
- Administrative changes are recorded with the actor, the action, the object, the IP address and the user agent.
- Payment notifications are recorded as described above.
- Model requests are recorded as metadata only — model, token counts, outcome, stop reason, latency and request identifier. Prompt text and model output are not columns in that record.
- Email sends and suppressions are recorded under idempotency keys.
- Edge and backend logs are forwarded to the telemetry provider named in the subprocessor table.
The applications you install
The macOS application and the command-line binaries are signed with the Company's Developer ID certificate, with a secure timestamp and the hardened runtime enabled, and are then submitted to Apple's notary service; the notarisation ticket is stapled to the application before the archive and the disk image are produced, so both carry it. The build refuses to present an artefact as notarised when no notary credential is configured, and says which step did not run. The Windows artefacts are signed through a managed code-signing service.
Release artefacts are published from the release pipeline in this repository rather than assembled by hand, and the version that ships is the version the pipeline produced.
Sub-processors, and the destinations you choose
The providers engaged to deliver the Services, what each handles, the data categories it receives and the region in which it operates are listed in the Privacy Policy. That table is the current list.
A model provider you configure with your own key or your own account is not a sub-processor of the Company. The product carries provider plugins for a long list of inference destinations — among them Anthropic, OpenAI, Google, Azure, Groq, OpenRouter, NVIDIA, Cerebras, Mistral, Cohere, Vertex, Bedrock and a generic OpenAI-compatible endpoint. Selecting one and supplying your credential makes the request yours: it is addressed to that provider, on your key, because you directed it, and that provider's terms and privacy notice govern it entirely. The Company does not stand between you and it and does not receive its content. Each plugin is used as provided and unmodified.
The one inference destination engaged by the Company is the hosted model service reached through the Company's own model proxy, which is named as a sub-processor in that table.
Framework
The control set is aligned to ISO/IEC 27001. The schema and the services are authored against it, and the alignment is asserted in the header of each database migration in this product. Access control, change management, logging, secret handling, supplier management and incident handling are organised against that framework.
This is an alignment, not a certification. The Company does not hold out a certification it has not been issued, and no statement on this page should be read as one.
Reporting a vulnerability
Report a suspected vulnerability by email to security@aivibe.in. Include what you found, the version and platform you found it on, and the steps that reproduce it. Do not include third-party data, and do not access, modify or retain any account that is not your own while investigating.
What a reporter can expect, as a matter of process:
- The report is acknowledged, and the acknowledgement states the next step in handling it.
- The reporter is kept informed of progress towards a fix, and may be asked for further information or for guidance on reproducing the finding.
- Where a report has not been acknowledged within six business days, it may be escalated to security@aivibe.in directly, or to support@aivibe.in.
- Good-faith research conducted within the scope below, and reported privately before disclosure, is not treated as a breach of the Terms of Use.
There is no bug bounty and no reward programme, and no timeline for a fix is offered. A finding is triaged on its merits.
Machine-generated reports are not accepted. A report produced by an automated scanner or by a language model, submitted without the reporter having reproduced and understood the finding, is closed unread.
The following are outside scope, for the reasons given above:
- Access to a server-mode instance that the operator chose to run without a password.
- Escaping the permission prompts, which are not a sandbox.
- How a model provider you configured handles data sent to it.
- The behaviour of an MCP server you connected.
- Modifying your own configuration files on your own machine.
- Findings that require physical access to an unlocked machine you already control.
Your part
Choose a strong and unique credential for the account you sign in with. Keep control of the device and the browser you sign in from, and sign out when you have finished. Read what the agent proposes before you approve it, and run it under an account with no more privilege than the work requires. Where the work demands true isolation, run it in a container or a virtual machine.
Governing law and jurisdiction
This document is governed by the laws of India. The courts at Chennai, Tamil Nadu, India have exclusive jurisdiction over every dispute arising out of or relating to it, wherever the party bringing it is situated and whatever the nature of the claim. This position applies equally to AIVEDHA INC, which operates under the parent's governance and jurisdiction.
Contact
AIVIBE SOFTWARE SERVICES PRIVATE LIMITED
A501 Saraduez Olympus, Chennai, Tamil Nadu 600100, India
Security: security@aivibe.in
General: support@aivibe.in
AIVEDHA INC
169 Madison Avenue, 62146, New York, NY 10016-5101
Email: support@aivedha.com