CVE-2026-57140 is a critical-severity (CVSS 9.4) Missing Authentication vulnerability in praisonai. A fix is available for praisonai — see the affected versions and patch details below.
PraisonAI AgentOS exposes unauthenticated agent listing and invocation
Exploitation Status
No confirmed exploitation observed yet
- CISA assesses this as automatable — exploitation doesn’t require manual, per-target effort, which raises the odds of mass scanning and opportunistic attacks.
- A successful exploit gives an attacker total control of the affected component, not partial access.
- CISA’s own triage has not observed active exploitation or public proof-of-concept code for this CVE as of its last assessment.
Exploitation and automatability from CISA’s SSVC triage for CVE-2026-57140.
EPSS Exploitation Probability
Probability of exploitation in the next 30 days, from FIRST.org EPSS.
How urgent is this, really
CVE-2026-57140 by exploitation likelihood (EPSS) against impact (CVSS). Outside the shaded patch-first corner.
Where this sits among everything scored
Of 386,684 CVEs with a current EPSS score, this one falls in the < 10% band (highlighted). Counts from FIRST.org, log-scaled.
Real-World Exposure
How broadly this vulnerability is actually deployed: weekly install volume shows current usage, and reverse-dependency count shows how many other packages break if it stays unpatched.
praisonainpmDescription
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/chatwithout 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/agentsdiscloses agent metadata and instruction prefixes; and/api/chatinvokesagent.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/agentsexposes instructions and/api/chatcan elicit data reachable by the agent.I:/api/chatlets unauthenticated callers drive agent/tool actions.A: unauthorized callers can consume model/API/server resources.
Suggested Fix
Recommended minimum fix:
- Add an authentication configuration to TypeScript
AgentOSConfig, for exampleauthToken,authRequired, or anauthorize(req)callback. - Default externally reachable servers to authenticated. Prefer fail-closed behavior when
hostis not loopback. - Change the default host from
0.0.0.0to127.0.0.1, matching the documented Python API-server hardening. - Register auth middleware before all non-health routes, including
/,/api/agents,/api/chat,/api/teams, and/api/flows. - Avoid returning agent instruction text from
/api/agentsunless the caller is authenticated and explicitly authorized. - Add regression tests that:
- unauthenticated
GET /api/agentsreturns401; - unauthenticated
POST /api/chatreturns401and does not callagent.chat; - authenticated requests still work;
- default
serve({ port })binds to loopback or fails closed when auth is not configured.
- unauthenticated
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, <= 1.7.1
Boundary:
<= 1.5.4 did not ship dist/os/agentos.js in the npm tarball.
No fixed npm version is known at the time of this report.
Version Sweep
The included sweep downloads npm tarballs and checks for the shipped dist/os implementation:
node poc/version_sweep_poc.js
Affected rows:
version has_agentos default_host_0_0_0_0 has_api_agents_instructions has_api_chat_agent_invocation has_401_unauthorized_guard mentions_authorization_header
1.6.0 true true true true false true
1.7.0 true true true true false true
1.7.1 true true true true false true
Earlier npm versions through 1.5.4 did not ship dist/os/agentos.js.
mentions_authorization_header is true because CORS allows the Authorization header. The sweep separately verifies there is no 401/Unauthorized route guard.
Advisory History
Checked:
- public GitHub advisories for
MervinPraison/PraisonAI; - private/triage advisories visible to this account; and
- visible PraisonAI advisories and prior reports.
No public or private advisory row in that data targets ecosystem: npm / package: praisonai.
Closest related advisories are Python/PyPI-scoped and do not cover the npm TypeScript package:
GHSA-pm96-6xpr-978x: PyPIpraisonai, unauthenticated information disclosure via Python AgentOS/api/agents, affected<= 4.5.120.GHSA-892r-p3jq-jp24: PyPIpraisonai, Python AgentOS unauthenticated remote agent invocation, affected>= 4.2.1, <= 4.6.57.GHSA-6rmh-7xcm-cpxj: PyPIpraisonai, generated legacy API server authentication disabled by default, affected>= 2.5.6, <= 4.6.33.GHSA-r7v3-x45f-g7hp/GHSA-7ww9-85pg-cv4x: PyPIpraisonai serve agents --api-keyignored.
This report should be tracked separately because it affects the npm package and the TypeScript implementation under src/praisonai-ts, with npm affected range >= 1.6.0, <= 1.7.1.
Affected Packages
| Ecosystem | Package | Vulnerable range | Fix |
|---|---|---|---|
| 📦npm | praisonai | ≥ 1.6.0&&< 1.7.2 | 1.7.2npm install praisonai@1.7.2 |
Detection & mitigation playbook
Open-source dependencyDetect
Scan your dependency tree (package-lock.json, pnpm-lock.yaml, requirements.txt, go.sum, etc.) for praisonai, including transitive dependencies — a direct dependency you never call can still pull in a vulnerable version.
Fix
Update praisonai to 1.7.2 or later, then make sure no transitive (indirect) dependency still pins the vulnerable range — O3 confirms CVE-2026-57140 is resolved across your whole dependency graph.
Workarounds
Put an independent control in front of the weakness: restrict the affected endpoint or interface to trusted networks, require an additional authentication factor or proxy-level check, and invalidate existing sessions and credentials in case the flaw has already been used.
Frequently Asked Questions
Is CVE-2026-57140 in your dependencies?
Find it across npm, including transitive dependencies.