Dev.to Security 🔐 Cybersecurity 👁 0 📖 3 min read

Onyx AI at 102 title matches: API keys connect users, agents, and enterprise data

Onyx AI at 102 title matches: API keys connect users, agents, and enterprise data Onyx is an open-source enterprise search and AI assistant that can connect to internal documents, applications, and chat workflows. A se

Onyx AI at 102 title matches: API keys connect users, agents, and enterprise data

Onyx is an open-source enterprise search and AI assistant that can connect to internal documents, applications, and chat workflows. A self-hosted instance is more than a search page. It can hold connectors, user groups, model-provider credentials, API keys, agent definitions, and actions that call other services. The external exposure question should follow those identities.

ZoomEye returned 102 matches for title="Onyx AI" and zero for app="Onyx AI" on November 18, 2026, using sub_type=all, page 1, and page size 1. The count is an index observation. It may include product pages and unrelated content, and it does not show whether any host is a live or misconfigured deployment.

Onyx's official developer page says the platform supports fine-grained API keys and APIs for messages, agents, LLMs, actions, and user authentication. Its product material describes permission-aware retrieval across connectors and support for self-hosted, cloud, or air-gapped deployment. Those features make identity integration central: the assistant has to decide which user can see which source, and which agent or API key can call which action.

Three identities in one deployment

The first identity is the human user. Group membership and source permissions determine which documents the assistant can retrieve. A public chat link or an unauthenticated search endpoint can bypass that model if it exposes results without the same authorization check.

The second identity is the API key. A key may be scoped to a user, a service, or an integration. The important question is whether it inherits the creator's access or has a narrower purpose. An API key that can create agents, add connectors, or inspect messages has more authority than a read-only chat token. Inventory keys by owner, scope, creation date, and last use.

The third identity is the agent or action. A connector may hold OAuth tokens for Drive, Slack, email, or another repository. An action may have access to a ticketing system or internal API. When a user asks the assistant to perform a task, the action should know whether it is acting for that user, as a shared service, or as an administrator. The presence of an LLM in the request path does not remove the need for object-level authorization.

What to inspect on an owned instance

Start from the application configuration and database, not from a title search. List users, groups, connectors, service accounts, API keys, agents, actions, and model providers. Check whether the deployment allows public signup or guest access. Review OAuth scopes and revoke connectors that no longer have an owner. Confirm that a low-privilege user cannot enumerate documents outside the connected source permissions.

Test the API separately from the browser. A request with no token should receive an authentication error, not a generic page that reveals deployment details. A token for one user should not read another user's chat. A connector credential should not be reusable as a general API token. If the application supports administrative actions, require a distinct role and log the change.

The language model provider adds another secret. A shared model key may be billed to the organization and may allow access to provider features that are not intended for every user. Use per-environment keys, spending limits, and rotation. Do not place the provider key in a prompt, a document, or a connector configuration that ordinary users can read.

Turning a broad count into a controlled check

The exact title phrase is useful because it is specific, but it still cannot distinguish a legitimate public product page from an internal assistant. Before requesting any response body, verify that the host belongs to the organization. Respect scope and do not attempt authentication against systems you do not own. For owned hosts, compare the public endpoint with the internal inventory and the expected deployment mode.

The result of the check should be a list of identities. Which human users can reach the assistant? Which keys can change its connectors? Which agents can call external tools? Which data sources trust the assistant as a user? A deployment that answers these questions has a meaningful security boundary around the product name.

One hundred two title matches are small enough to review manually but large enough to show why automated classification can mislead. The product's value comes from joining systems together. Its risk comes from allowing one key or one agent to join them too broadly.

Query evidence

title="Onyx AI"
sub_type=all, page=1, pagesize=1, observed 2026-11-18T07:42:27+08:00, count=102

References

📰 Read the original article on Dev.to Security

Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.