VulnWatch VulnWatch
← Back to dashboard
Critical github · GHSA-9752-mhqh-h34f

npm PraisonAI AgentOS exposes unauthenticated agent listing and invocation

Published Jun 18, 2026 CVSS 9.4

Summary

The published npm package praisonai ships a TypeScript AgentOS HTTP server that defaults to host: "0.0.0.0" and registers sensitive agent routes without any authentication or authorization middleware.

When a developer starts AgentOS, a network attacker who can reach the service can:

  • read configured agent names, roles, and the first 100 characters of each agent's instructions through GET /api/agents; and
  • invoke the selected agent through POST /api/chat without credentials.

This is distinct from the existing Python/PyPI AgentOS and API-server advisories. The affected package here is npm:praisonai; the current published npm package is 1.7.1, and the same TypeScript source is still present in refreshed origin/main at v4.6.58.

Technical Details

AgentOSConfig exposes host, CORS, and API-prefix settings but no authentication token, auth mode, or authorization callback.

Relevant current-head source:

src/praisonai-ts/src/os/config.ts
  26: host?: string;              // default: "0.0.0.0"
  35: corsOrigins?: string[];     // default: ["*"]
  66: export const DEFAULT_AGENTOS_CONFIG = {
  68:     host: '0.0.0.0',
  71:     corsOrigins: ['*'],

AgentOS._createApp() registers JSON parsing and CORS handling, then immediately registers routes. There is no middleware between body parsing and route registration that validates an API key, bearer token, session, origin-bound secret, or any other credential.

Relevant current-head source:

src/praisonai-ts/src/os/agentos.ts
  179: app.use(express.json());
  182: // Add CORS middleware
  204: // Register routes
  205: this._registerRoutes(app);

The sensitive routes are then exposed:

src/praisonai-ts/src/os/agentos.ts
  235: app.get(`${apiPrefix}/agents`, ...)
  240: instructions: agent.instructions ? ... : null
  250: app.post(`${apiPrefix}/chat`, ...)
  273: const response = await agent.chat(message);
  331: const host = options.host || this.config.host;
  338: this._server = app.listen(port, host, ...)

Because the default host is 0.0.0.0, await app.serve({ port: 8000 }) listens on all interfaces unless the developer explicitly overrides host.

Why This Is Not Intended Behavior

PraisonAI's official TypeScript documentation describes the npm package as a production-ready multi-agent framework and directs users to install it with npm install praisonai.

PraisonAI's security documentation says security reports should include affected versions, impact, reproduction steps, and a suggested fix, and states that GitHub Security Advisories are the preferred reporting method. The same security page also documents a prior hardening change where API servers were changed to require authentication by default and bind to 127.0.0.1 instead of 0.0.0.0.

The TypeScript npm AgentOS implementation still does the opposite:

  • default bind address is 0.0.0.0;
  • no auth config exists in AgentOSConfig;
  • /api/agents discloses agent metadata and instruction prefixes; and
  • /api/chat invokes agent.chat(message) directly.

The patched-control branch in the PoV confirms that adding a pre-route bearer-token middleware makes the same unauthenticated requests fail with 401.

PoV

The PoV installs the published npm package in a temporary project, starts AgentOS on 127.0.0.1 with a mock agent, and sends loopback HTTP requests. It does not call any LLM provider or external service after package installation.

Run from a local reproduction checkout:

node poc/pov_poc.js 1.7.1

Observed result:

{
  "version": "1.7.1",
  "defaultHost": "0.0.0.0",
  "agentsStatus": 200,
  "agentsBody": {
    "agents": [
      {
        "name": "finance-admin",
        "role": "internal finance operations",
        "instructions": "poc SECRET: refund-wire-tool may alter customer balances"
      }
    ]
  },
  "chatStatus": 200,
  "chatBody": {
    "response": "agent-invoked:transfer-check",
    "agent_name": "finance-admin"
  },
  "invokedMessages": [
    "transfer-check"
  ]
}

No Authorization header is sent in the vulnerable requests.

The PoV also applies a minimal local-only auth middleware patch to the temporary installed copy and reruns the same requests as a control:

{
  "patchedNoAuthAgents": 401,
  "patchedNoAuthChat": 401,
  "patchedWithAuthAgents": 200,
  "patchedWithAuthChat": 200
}

This control demonstrates that the PoV is exercising the missing authentication boundary, not an artifact of the mock agent.

PoC

The PoV section above contains the local reproduction command, input, and decisive output.

Impact

An attacker who can reach a running TypeScript AgentOS server can invoke configured agents without credentials. Real impact depends on the deployed agent, but PraisonAI agents may have access to tools, memory, workflow state, external APIs, credentials in process environment, and business data. Unauthorized prompt injection through /api/chat can therefore affect confidentiality and integrity of downstream systems reachable by the configured agent.

GET /api/agents also discloses agent names, roles, and instruction prefixes, which can reveal internal workflow details and help tailor prompts against the exposed agent.

This report does not claim arbitrary code execution by default. If the deployed agent has code, file, browser, MCP, or business-operation tools, the unauthenticated invocation endpoint can become the entry point for those tool-side effects.

Severity

Suggested severity: Critical.

Rationale:

  • AV: the vulnerable component is an HTTP service and defaults to all-interface binding.
  • AC: exploitation is a direct HTTP request.
  • PR: no credentials are required.
  • UI: no user interaction is required after the server is running.
  • S: impact is within the vulnerable service and the configured agent's authority.
  • C: /api/agents exposes instructions and /api/chat can elicit data reachable by the agent.
  • I: /api/chat lets unauthenticated callers drive agent/tool actions.
  • A: unauthorized callers can consume model/API/server resources.

Suggested Fix

Recommended minimum fix:

  1. Add an authentication configuration to TypeScript AgentOSConfig, for example authToken, authRequired, or an authorize(req) callback.
  2. Default externally reachable servers to authenticated. Prefer fail-closed behavior when host is not loopback.
  3. Change the default host from 0.0.0.0 to 127.0.0.1, matching the documented Python API-server hardening.
  4. Register auth middleware before all non-health routes, including /, /api/agents, /api/chat, /api/teams, and /api/flows.
  5. Avoid returning agent instruction text from /api/agents unless the caller is authenticated and explicitly authorized.
  6. Add regression tests that:
    • unauthenticated GET /api/agents returns 401;
    • unauthenticated POST /api/chat returns 401 and does not call agent.chat;
    • authenticated requests still work;
    • default serve({ port }) binds to loopback or fails closed when auth is not configured.

Affected Package/Versions

  • Repository: MervinPraison/PraisonAI
  • Ecosystem: npm
  • Package: praisonai
  • Current npm version: 1.7.1
  • Component: src/praisonai-ts/src/os/agentos.ts
  • Config component: src/praisonai-ts/src/os/config.ts
  • Refreshed repo head checked: 1ad58ca02975ff1398efeda694ea2ab78f20cf3e (v4.6.58)

Confirmed affected npm versions:

>= 1.6.0,

Affected AI Products

prompt injection llm
Get the weekly digest. Every Monday: top AI security stories of the week. Free.