Skip to content

CVE-2026-111935: NoteGen Multimodal AI Chat Image Analysis Local File Exfiltration & Companion AI / Streaming Data-Plane Flaws in RisingWave, SGLang & Television (CVE-2026-112262, CVE-2026-112082, CVE-2026-112263)

Density:
C-Level Executive Briefing CVE-2026-111935 CRITICAL SCORE 92/100

Critical agentic AI, streaming database, Mixture-of-Experts (MoE) LLM scheduler, and developer CLI vulnerability cluster disclosed on October 19-23, 2026. Under CVE-2026-111935 (CWE-73), NoteGen versions through 0.38.3 trust synced sourceUrl file paths during multimodal chat image analysis (image_inspect), allowing attackers who poison a shared Git/WebDAV/S3 sync repository to force the victim's client to read arbitrary local files and upload them to an external AI vision endpoint. Dissected alongside RisingWave < 3.1.0 missing authorization on posix_fs and fs SQL connectors (CVE-2026-112262, allowing any SQL user with CREATE schema rights to read and write arbitrary host files), SGLang <= 0.5.19 Scheduler.expert_distribution_handle DoS in python/sglang/srt/eplb/expert_distribution.py (CVE-2026-112082), and Television (tv) 0.15.0-0.15.9 OS command injection via single-quote breakout in format_command (CVE-2026-112263).

HYBRID

HERMES THREAT SCORE & MULTIMODAL AI / STREAMING DATA-PLANE EVALUATION

Target: NoteGen image_inspect Synced Attachment Loader (<= 0.38.3), RisingWave posix_fs/fs SQL Connectors (< 3.1.0) & SGLang EPLB Scheduler
Confidence: 98%
92 / 100
CRITICAL

Measures real-world operational relevance, exploit weaponization, and active threat posture.

Dimension Breakdown
Exploitability 19 / 20
Threat Activity 17 / 20
Weaponization 19 / 20
Exposure 18 / 20
Prevalence 18 / 20
Impact 19 / 20
Exploit Maturity 19 / 20
Attack Chain Potential 19 / 20
โš–๏ธ Divergence & Operational Rationale

While individual CVSS v3.1 scores rate NoteGen CVE-2026-111935, RisingWave CVE-2026-112262, SGLang CVE-2026-112082, and Television CVE-2026-112263 between 7.5 and 8.8, the Hermes Threat Score evaluates this cluster at 92 (CRITICAL) with a Hermes Agentic Security Score (HASS) of 94 (CRITICAL). First, NoteGen (<= 0.38.3) exemplifies a high-impact cross-boundary agentic attack vector: by poisoning `sourceUrl` inside a shared Git/WebDAV/S3 note sync repository with an absolute host path or `file://` URI, an attacker causes the victim's `image_inspect` pipeline to read sensitive local files and transmit their contents out-of-band to a third-party multimodal LLM API. Second, RisingWave (< 3.1.0, CVE-2026-112262) powers real-time streaming RAG and feature pipelines where LLM text-to-SQL agents or low-privilege analysts hold `CREATE` privileges on a schema; omitting superuser checks on `connector = 'posix_fs'` and `connector = 'fs'` turns any SQL connection into arbitrary host file read (`/etc/shadow`, cloud credentials) and arbitrary file write.

๐Ÿ•ธ๏ธ Connected Knowledge Graph & Provenance

CVE-2026-111935: NoteGen Multimodal AI Chat Image Analysis Local File Exfiltration & Companion AI / Streaming Data-Plane Flaws in RisingWave, SGLang & Television (CVE-2026-112262, CVE-2026-112082, CVE-2026-112263)VULNERABILITY

Connected Nodes: 1
Active Relationships (Outgoing)
→ affectsPRODUCTNoteGen Multimodal AI Note-Taking Assistant
98% VERY_HIGH

Software platform affected by security vulnerabilities and agentic attack patterns.

๐Ÿ” Why is this related? (Evidence & Provenance)

“Confirmed security vulnerability in NoteGen Multimodal AI Note-Taking Assistant documented in Hermes dossier.”

Supporting Verified Evidence:

Between October 19 and October 23, 2026, VulnCheck Initial Access Intelligence and the National Vulnerability Database (NVD) published weaponized advisories for four critical flaws spanning multimodal AI note-taking assistants (NoteGen), distributed SQL streaming databases (RisingWave), Mixture-of-Experts (MoE) LLM inference engines (SGLang), and developer terminal fuzzy finders (Television):

  1. CVE-2026-111935 (CWE-73 / CWE-200, NoteGen <= 0.38.3 Local File Disclosure via Chat Image Analysis): In NoteGen (versions through 0.38.3), the multimodal AI chat assistant includes an image_inspect feature that loads attachment files referenced by sourceUrl in synced note records. Because NoteGen fails to validate or confine sourceUrl to the applicationโ€™s workspace before reading the file from disk, an attacker with write access to a shared sync target (Git, WebDAV, or S3) can inject an absolute filesystem path (/Users/<victim>/.ssh/id_rsa, C:\Users\<victim>\.aws\credentials, or local screenshots) or file:// URL, forcing the victimโ€™s client to read the local file and transmit it to the configured AI provider.
  2. CVE-2026-112262 (CWE-862, RisingWave < 3.1.0 Missing Authorization via posix_fs and fs Connectors): In RisingWave versions prior to 3.1.0, the SQL DDL handlers for CREATE SOURCE ... WITH (connector = 'posix_fs', ...) and CREATE SINK ... WITH (connector = 'fs', ...) require only standard schema CREATE privileges rather than superuser/admin authorization. Any low-privilege SQL user (or compromised text-to-SQL AI agent) can read arbitrary server files (match_pattern = '/etc/shadow') and write arbitrary files to the host filesystem with the privileges of the risingwave server process (often root in containerized deployments).
  3. CVE-2026-112082 (CWE-400, SGLang <= 0.5.19 Denial of Service via Scheduler.expert_distribution_handle): Complements our recent analysis of the SGLang 0.5.22 Disaggregated Decode SSRF & Mooncake KV-Cache Exhaustion Chain (CVE-2026-112267). In python/sglang/srt/eplb/expert_distribution.py, the Expert-Parallelism Load Balancer (EPLB) recording handler Scheduler.expert_distribution_handle in SGLang <= 0.5.19 can be triggered remotely to induce a Denial of Service (DoS) across Mixture-of-Experts (MoE) GPU inference workers.
  4. CVE-2026-112263 (CWE-78, Television 0.15.0โ€“0.15.9 OS Command Injection via External Action Formatting): In the Rust terminal fuzzy finder Television (tv) versions 0.15.0 through 0.15.9, the format_command helper improperly escapes single quotes when interpolating selected file entries into external shell action templates. When a developer or coding agent triggers an external action on a maliciously named file inside an untrusted repository, the unescaped single quote breaks out of the quoted argument and executes arbitrary shell commands.

1. CVE-2026-111935: Unvalidated sourceUrl Path Resolution in NoteGen image_inspect (<= 0.38.3)

Section titled โ€œ1. CVE-2026-111935: Unvalidated sourceUrl Path Resolution in NoteGen image_inspect (<= 0.38.3)โ€

NoteGen allows users to synchronize Markdown notes, marks, and chat context across devices or collaborative teams using Git repositories, WebDAV shares, or S3-compatible buckets. When a user invokes multimodal chat analysis on a note containing image marks, the client resolves each recordโ€™s sourceUrl and reads the file bytes to construct a base64 multimodal vision payload (image_url: { url: "data:image/png;base64,..." }):

// VULNERABLE PATTERN: NoteGen Chat Image Inspection Handler (<= 0.38.3)
async function inspectSyncedImage(markRecord: SyncedMarkRecord, aiClient: OpenAIClient) {
const rawPath = markRecord.sourceUrl; // Synced from Git / WebDAV / S3 JSON metadata!
// FLAW (CWE-73): Accepts arbitrary absolute paths (/etc/..., /Users/..., C:\...)
// or file:// URIs from synced metadata without verifying containment inside the
// NoteGen appData / workspace directory!
const normalizedPath = rawPath.startsWith('file://')
? fileURLToPath(rawPath)
: rawPath;
const fileBytes = await readFile(normalizedPath);
const base64Data = Buffer.from(fileBytes).toString('base64');
// Transmits the local file contents directly to the AI vision endpoint / chat log!
return await aiClient.chat.completions.create({
model: markRecord.model || 'gpt-4o',
messages: [
{
role: 'user',
content: [
{ type: 'text', text: 'Analyze and transcribe all text inside this image:' },
{ type: 'image_url', image_url: { url: `data:image/png;base64,${base64Data}` } }
]
}
]
});
}

If an attacker with write access to the sync target modifies a synced markโ€™s sourceUrl to point to a local file on the victimโ€™s machine (and optionally points the workspaceโ€™s custom OpenAI-compatible base URL or reads the synced chat transcript), NoteGen reads the victimโ€™s local file during image_inspect and exfiltrates its contents into the AI conversation history.

2. CVE-2026-112262: RisingWave < 3.1.0 posix_fs Source & fs Sink Arbitrary File Read/Write

Section titled โ€œ2. CVE-2026-112262: RisingWave < 3.1.0 posix_fs Source & fs Sink Arbitrary File Read/Writeโ€

RisingWave provides built-in posix_fs (local POSIX filesystem source) and fs (local filesystem sink) connectors primarily intended for local testing or privileged batch ingestion. However, prior to version 3.1.0, RisingWave did not gate posix_fs or fs connector instantiation behind superuser privileges or a cluster-level configuration allowlist:

-- VULNERABLE EXPLOITATION PATTERN: RisingWave (< 3.1.0) CVE-2026-112262
-- Executed by any low-privilege SQL user holding CREATE privilege on a schema:
-- 1. Arbitrary Host File Read via posix_fs Connector:
CREATE SOURCE host_secret_exfil (line VARCHAR) WITH (
connector = 'posix_fs',
posix_fs.root = '/etc',
match_pattern = 'shadow*'
) FORMAT PLAIN ENCODE TEXT;
-- 2. Arbitrary Host File Write (Cron / SSH Authorized Keys RCE) via fs Sink:
CREATE SINK host_cron_rce FROM malicious_payload_table WITH (
connector = 'fs',
fs.path = '/var/spool/cron/crontabs',
type = 'append-only'
);

Because the RisingWave compute node executes the connector I/O directly on the host/container filesystem, any tenant or SQL agent tool can read /etc/shadow, /proc/self/environ (leaking AWS/GCP object-store keys), or write cron jobs and SSH authorized_keys.

3. CVE-2026-112082 (SGLang EPLB DoS) & CVE-2026-112263 (Television format_command OS Command Injection)

Section titled โ€œ3. CVE-2026-112082 (SGLang EPLB DoS) & CVE-2026-112263 (Television format_command OS Command Injection)โ€
  • CVE-2026-112082 (SGLang <= 0.5.19): In python/sglang/srt/eplb/expert_distribution.py, the Scheduler.expert_distribution_handle method manages Expert-Parallelism Load Balancing (EPLB) telemetry for Mixture-of-Experts (MoE) models (such as DeepSeek-V3/R1 and Mixtral). Unauthenticated or unvalidated invocation of the expert distribution recording state transitions causes unhandled scheduler exceptions or resource exhaustion.
  • CVE-2026-112263 (Television 0.15.0โ€“0.15.9): When formatting external action commands (format_command), Television wraps the selected entry in single quotes ('{}') without escaping embedded single quotes (' -> '\''). Selecting a file named poc'; id > /tmp/pwned; '.txt executes arbitrary shell commands immediately.
flowchart TD
subgraph NoteGen["CVE-2026-111935: NoteGen (<= 0.38.3)"]
NG1["Shared Git / WebDAV / S3 Sync Target<br/>Poisoned Mark Metadata: sourceUrl = 'file:///Users/victim/...'"]
NG2["CWE-73: Unvalidated Path in image_inspect<br/>Reads Local Host File Outside Workspace"]
NG3["Transmits Local File via Multimodal Vision API<br/>& Synced Chat Transcript Exfiltration"]
NG1 --> NG2 --> NG3
end
subgraph DataAndAI["Companion AI, Streaming SQL & CLI Advisories"]
RW["CVE-2026-112262: RisingWave (< 3.1.0)<br/>CREATE SOURCE (connector='posix_fs') & CREATE SINK (connector='fs')<br/>Low-Privilege SQL User Arbitrary Host File Read/Write"]
SG["CVE-2026-112082: SGLang (<= 0.5.19)<br/>Scheduler.expert_distribution_handle (expert_distribution.py)<br/>MoE Expert-Parallelism Load Balancer DoS"]
TV["CVE-2026-112263: Television TUI (0.15.0 - 0.15.9)<br/>format_command Single-Quote Breakout<br/>Malicious Filename OS Command Injection"]
end
style NG2 fill:#ff1a1a,stroke:#800000,color:#ffffff
style NG3 fill:#ff4d4d,stroke:#990000,color:#ffffff
style RW fill:#ff1a1a,stroke:#800000,color:#ffffff
style SG fill:#ff6600,stroke:#993300,color:#ffffff
style TV fill:#ff6600,stroke:#993300,color:#ffffff

The following script audits synced NoteGen workspace repositories for suspicious out-of-workspace sourceUrl paths (CVE-2026-111935) and checks Git repositories for single-quote command injection filenames targeting Television (CVE-2026-112263):

#!/usr/bin/env python3
"""
Defensive Audit Script:
1. CVE-2026-111935 โ€” Scan NoteGen synced JSON/Markdown files for external sourceUrl paths
2. CVE-2026-112263 โ€” Scan repository filenames for single-quote shell breakout patterns
"""
import os
import re
from pathlib import Path
WORKSPACE = Path(os.environ.get("AUDIT_WORKSPACE", "."))
# 1. Audit for CVE-2026-111935 suspicious sourceUrl values
source_url_re = re.compile(r'"sourceUrl"\s*:\s*"([^"]+)"')
for p in WORKSPACE.rglob("*.json"):
try:
text = p.read_text(errors="ignore")
for match in source_url_re.finditer(text):
val = match.group(1)
if val.startswith(("file://", "/", "C:\\", "\\\\")) or ".." in val:
print(f"[!] CVE-2026-111935 suspicious sourceUrl in {p}: {val}")
except Exception:
pass
# 2. Audit for CVE-2026-112263 malicious single-quote filenames
for p in WORKSPACE.rglob("*"):
if "'" in p.name and any(sep in p.name for sep in (";", "|", "&", "$(")):
print(f"[!] CVE-2026-112263 suspicious filename detected: {p}")

title: RisingWave posix_fs/fs Arbitrary File Access, SGLang EPLB DoS & Television Command Injection (CVE-2026-112262 / CVE-2026-112082 / CVE-2026-112263)
id: 111935b4-2026-4e77-9c12-aidataplane23
status: experimental
description: Detects exploitation of RisingWave posix_fs and fs SQL connectors for host file read/write (CVE-2026-112262), SGLang expert_distribution_handle abuse (CVE-2026-112082), and NoteGen / Television local file and command injection patterns (CVE-2026-111935 / CVE-2026-112263).
references:
- https://vulncheck.com/advisories/notegen-through-0.38.3-local-file-disclosure-via-chat-image-analysis
- https://vulncheck.com/advisories/risingwave-before-3.1.0-missing-authorization-via-posix-fs-and-fs-connectors
- https://nvd.nist.gov/vuln/detail/CVE-2026-112082
- https://vulncheck.com/advisories/television-0.15.0-through-0.15.9-os-command-injection-via-external-action-formatting
author: Hermes Threat Intelligence
date: 2026/10/23
logsource:
category: application
detection:
selection_risingwave_fs_connectors:
query|re: '(?i)CREATE\s+(SOURCE|TABLE|SINK)\s+.*connector\s*=\s*''(posix_fs|fs)'''
selection_sglang_eplb:
message|contains:
- 'expert_distribution_handle'
- 'start_expert_distribution_record'
- 'dump_expert_distribution_record'
condition: selection_risingwave_fs_connectors or selection_sglang_eplb
falsepositives:
- Authorized local development testing of RisingWave posix_fs sources
level: high
tags:
- attack.initial_access
- attack.execution
- attack.collection
- attack.t1005
- attack.t1059
- cve.2026-111935
- cve.2026-112262
- cve.2026-112082
- cve.2026-112263

  1. Upgrade NoteGen Beyond 0.38.3 (>= 0.38.4, CVE-2026-111935): Update NoteGen to version 0.38.4 or later, which validates sourceUrl paths during image_inspect to reject file:// URIs, directory traversal (..), and absolute paths outside the local NoteGen workspace directory. Audit shared Git, WebDAV, and S3 synchronization repositories for modified note metadata.
  2. Upgrade RisingWave to >= 3.1.0 (CVE-2026-112262): Update RisingWave clusters to 3.1.0 or later, which restricts posix_fs and fs connectors to authorized administrators/configurations. Until upgraded, revoke CREATE privileges on schemas from untrusted SQL users or LLM text-to-SQL service accounts, and inspect rw_catalog.rw_sources and rw_catalog.rw_sinks for any posix_fs or fs definitions.
  3. Upgrade SGLang to >= 0.5.23 (CVE-2026-112082 & CVE-2026-112267) and Television (tv) to >= 0.15.10 (CVE-2026-112263): Update SGLang to 0.5.23 or later (remediating both Scheduler.expert_distribution_handle DoS CVE-2026-112082 and the 0.5.22 disaggregated Decode/Mooncake vulnerabilities CVE-2026-112266โ€“112270), and upgrade the television (tv) binary to 0.15.10 or later so format_command properly escapes single quotes.