Summary
PraisonAI exposes sensitive agent configuration and chat invocation APIs by default without requiring authentication or scoped network binding.
When deployed as-is, unauthenticated remote users can read agent instructions and system prompts or trigger agent actions directly via the public API.
Affected context
This vulnerability appears in PraisonAI deployments that run with the default settings prior to version 1.7.3.
In this configuration, the service binds to all network interfaces, does not enforce API keys, and allows wildcard CORS, making the REST endpoints globally reachable and callable by any origin when placed on a routable host or behind permissive reverse proxies.
Exploit preconditions
- Attack surface: HTTP API endpoints exposed on
/api/agentsand/api/chat. - Required access: Network reachability to the PraisonAI instance (e.g., via LAN, VPN, or public Internet).
- Trigger point: Direct GET or POST requests to the affected endpoints from any client without authentication.
- Why exploitation is possible: The default configuration trusts all origins, binds to all interfaces, and omits any API key or authentication requirement, turning private operational APIs into publicly accessible resources.
Why this happens
The core issue is an insecure-by-default configuration that fails to treat operational AI control and configuration endpoints as sensitive resources.
Binding to all interfaces and omitting API key enforcement means the service does not differentiate between trusted and untrusted clients.
Combined with wildcard CORS headers, browsers and other clients are permitted to interact with the API without restriction.
This maps to CWE-200: Exposure of Sensitive Information to an Unauthorized Actor, where sensitive operational data and capabilities are made available due to overly permissive defaults and missing access controls.
Potential impact
- Unauthenticated users may read agent instructions, system prompts, and configuration metadata.
- Attackers could invoke agents via
/api/chat, potentially driving unintended actions or workflows. - Internal business logic or proprietary prompt content could be exposed to external parties.
- Misuse of agents may lead to data leakage, reputational damage, or abuse of integrated downstream services.
- The instance can be used as a free, unmetered AI backend by unauthorized users, consuming resources and affecting availability.
General mitigation guidance
Mitigation centers on making operational AI APIs secure-by-default and tightening the trust boundary around configuration and chat endpoints.
Defensive patterns typically include:
- Requiring strong authentication (API keys or tokens) for all non-public endpoints and enforcing authorization checks server-side.
- Binding the service to specific interfaces (e.g., localhost or internal addresses) unless explicit exposure is required, and placing it behind controlled gateways.
- Replacing wildcard CORS with explicit allowlists for trusted origins and disallowing cross-origin access where it is not needed.
- Treating agent instructions, system prompts, and chat endpoints as sensitive controls and applying least privilege to any clients interacting with them.
- Providing hardening presets or secure configuration profiles that are enabled by default, so operators must explicitly opt in to relaxed behaviors.
How teams detect issues like this
Detection typically involves both configuration review and runtime monitoring:
- Reviewing deployment manifests, environment variables, and configuration files for interface binding, CORS settings, and missing API key or auth parameters.
- Using static analysis or infrastructure-as-code linters to flag services that bind to
0.0.0.0without accompanying auth controls or that set wildcard CORS headers. - Implementing policy-as-code (e.g., in CI/CD) that enforces authentication and origin restrictions on AI-related control APIs.
- Monitoring access logs for unauthenticated calls to
/api/agentsand/api/chat, especially from external networks or unexpected origins. - Applying automated security checks that verify that sensitive endpoints require credentials before deployment.

