# Microsoft/GitHub AI Pilot Reports 99 Verified Security Issues in Express, ImageMagick, Caddy, AutoGPT, FastMCP and Netty

> Microsoft and GitHub sent 99 verified private security reports to six open-source projects after an AI-assisted scanning pilot (Codename MDASH using the MAI Cyber model) flagged 9,791 possible issues. As of 2026-11-15 maintainers had fixed 55 of the reports, but only 14 of those fixes had shipped in releases, so users of the affected projects may still run unpatched code.

- **Published:** 2026-11-18T23:01:24Z
- **Last reviewed:** 2026-11-18T23:01:24Z
- **Canonical:** https://intel.threadlinqs.com/threat/TL-2026-6387
- **Card image:** https://intel.threadlinqs.com/og/threat/TL-2026-6387.png
- **ID:** TL-2026-6387
- **Severity:** INFO
- **Category:** VULNERABILITY
- **Status:** MONITORING
- **Detections:** 9 (1 shown free below) · **IOCs:** 12

## Description

On 2026-11-17 GitHub (Ashley Wolf, Gregg Cochran) published results of a pilot in which Microsoft's Codename MDASH, a multi-model agentic scanning harness, ran an earlier version of the MAI Cyber model against six open-source projects: AutoGPT, Caddy, Express, FastMCP, ImageMagick and Netty. All six are GitHub Secure Open Source Fund alumni, and maintainers approved the scans beforehand. The scan flagged 9,791 possible issues. A GitHub Security Lab AI triage agent, built on tailored Copilot skills, removed duplicates, checked validity and prioritized higher-impact bugs. Seven Microsoft FORGE Lab researchers then reviewed the remaining findings by hand, about two days per project. That left 99 verified reports, roughly 1 per 100 raw findings.

GitHub delivered the 99 reports privately through GitHub private vulnerability reporting. Each report carried a proof of concept and a suggested fix. Maintainers decided what to fix, how and when. As of 2026-11-15, 55 reports were fixed but only 14 of those fixes were in a release. Per-project figures are published only for Caddy (22 reports, all fixed within a week or less) and ImageMagick (22 reports, handled by a two-person team, one fix pushed in 15 minutes). For AutoGPT, a report on a recently shipped security check arrived three days after it shipped; maintainers wrote a fix within a day and had it in production in under a week. Express, FastMCP and Netty have no published counts. The ImageMagick findings are separate from previously reported ImageMagick memory-corruption flaws.

The sources name no CVE or GHSA identifiers, no CVSS scores, no vulnerability classes for the 99 reports, no exploitation in the wild and no threat actor. This record therefore documents an exposure and patch-latency situation, not an observed attack. Because the fixed-versus-released gap (55 vs 14) means some fixes exist only in source, defenders should track upstream releases and advisories for these projects and update to releases that contain the fixes. Netty's maintainer said finding vulnerabilities is no longer the bottleneck and fixing them is. GitHub also notes open gaps: delivering fixes privately, CI/CD testing on private forks, maintainer access to cyber-capable models to verify proofs of concept, and manual review that does not yet scale.

MDASH was introduced by Microsoft on 2026-05-12 as a system of more than 100 specialized AI agents (Prepare, Scan, Validate, Dedupe, Prove stages) in limited private preview. Microsoft credited it with 16 Windows CVEs in May 2026. Microsoft announced MAI Cyber 2 on the same day as the GitHub post. The pilot covered only six projects and is an early test, not proof the approach scales.

## MITRE ATT&CK

- T1190 Exploit Public-Facing Application
- T1195.001 Compromise Software Dependencies and Development Tools
- T1203 Exploitation for Client Execution
- T1587.004 Develop Capabilities: Exploits
- T1588.006 Obtain Capabilities: Vulnerabilities

## Free detection (Splunk SPL, Microsoft KQL, Sigma)

**ImageMagick Invoked With Dangerous Coder or Remote-Fetch Arguments (Exploitation)** (`TL-2026-6387-SPL-001`)

Behavioural detection for ImageMagick (imagemagick) invoked with pseudo-protocol coders (msl:, ephemeral:, url:, http:, mvg:, ps:, pdf:) or @file indirection. ImageMagick received 22 of the 99 verified MDASH pilot reports (TL-2026-6387) and some fixes are not yet in releases; hosts that pass untrusted files to ImageMagick are the exposure. Maps to T1190 and T1203; T1059 covers delegate-driven command execution. No CVE or exploit signature is published, so this is behaviour-based.

- **Languages:** Splunk SPL, Microsoft KQL, Sigma
- **ATT&CK:** T1190, T1203, T1059
- **Data sources:** sysmon, sysmon_for_linux, linux_edr_process
- **Severity:** medium
- **Confidence:** medium

### Splunk SPL

```spl
index=endpoint sourcetype="XmlWinEventLog:Microsoft-Windows-Sysmon/Operational" EventCode=1 (Image="*/magick" OR Image="*magick.exe" OR Image="*/convert" OR Image="*convert.exe" OR Image="*/mogrify" OR Image="*/identify" OR Image="*identify.exe")
| where match(CommandLine, "(?i)(^|\\s)(msl|ephemeral|url|https?|ftp|mvg|ps|pdf|inline):|(label|text):@|\\s@/")
| stats count min(_time) as first_seen max(_time) as last_seen values(CommandLine) as command_lines values(ParentImage) as parent_images by host, User, Image
| convert ctime(first_seen) ctime(last_seen)
```

### Microsoft KQL

```kql
DeviceProcessEvents
| where Timestamp > ago(24h)
| where FileName in~ ("magick","magick.exe","convert","convert.exe","mogrify","identify","identify.exe")
| where ProcessCommandLine matches regex @"(?i)(^|\s)(msl|ephemeral|url|https?|ftp|mvg|ps|pdf|inline):|(label|text):@|\s@/"
| project Timestamp, DeviceName, AccountName, FileName, ProcessCommandLine, InitiatingProcessFileName, InitiatingProcessCommandLine
```

### Sigma

```yaml
title: ImageMagick Invoked With Dangerous Coder or Remote-Fetch Arguments
id: 6f1c2a8e-3b47-4d19-9a52-7c0e5d8b1a01
status: experimental
description: Detects ImageMagick binaries launched with pseudo-protocol coders (msl, ephemeral, url, http, mvg, ps, pdf) or @file indirection, the pattern used when untrusted input reaches ImageMagick. Relevant to the 22 verified ImageMagick reports in the MDASH pilot; no CVE is published.
references:
    - https://github.blog/open-source/maintainers/maintainers-dont-need-more-ai-bug-reports-they-need-fixes/
author: Threadlinqs Team
date: 2026/11/18
tags:
    - attack.initial-access
    - attack.t1190
    - attack.execution
    - attack.t1203
    - attack.t1059
logsource:
    category: process_creation
    product: linux
detection:
    selection_img:
        Image|endswith:
            - '/magick'
            - '/convert'
            - '/mogrify'
            - '/identify'
    selection_cli:
        CommandLine|re: '(?i)(^|\s)(msl|ephemeral|url|https?|ftp|mvg|ps|pdf|inline):|(label|text):@|\s@/'
    condition: all of selection_*
falsepositives:
    - Image pipelines that intentionally fetch remote images with https: or url: coders
    - Administrators running ImageMagick manually for PDF or PostScript conversion
level: medium

```

**Known false positives:**

- Image pipelines that deliberately fetch remote images via https: or url: coders
- Administrators converting PDF or PostScript files manually with ImageMagick

**Tuning notes:** Allowlist known image-processing service accounts and hosts. Narrow Image to the ImageMagick paths actually deployed. Adjust the coder regex to your upload workflow; raise severity when the parent process is a web runtime (node, java, php-fpm, python).

### 8 more detection rules (Splunk SPL, Microsoft KQL, Sigma), unlock with TL Creds

- Web-Facing Runtime Spawning Shell (Exploitation) (locked: unlock with TL Creds)
- Web Runtime Writing Persistence Locations (Post-Exploitation) (locked: unlock with TL Creds)
- Web Runtime Spawning Download or Remote-Access Tools (C2 / Lateral Movement) (locked: unlock with TL Creds)
- Affected Project Installed From Non-Default Package Registry (Supply Chain) (locked: unlock with TL Creds)
- Repeated Segfaults in [indicator], [indicator], Node or Java (Exploit Attempt) (locked: unlock with TL Creds)
- Runtime or [indicator] Access to Cloud Metadata Service (Credential Theft / SSRF) (locked: unlock with TL Creds)
- Web Runtime or [indicator] Writing Webshell or Payload Files (Post-Exploitation) (locked: unlock with TL Creds)
- [indicator] or Web Runtime Outbound to Public IP on Unusual Port (C2 / SSRF) (locked: unlock with TL Creds)

Sign up free to unlock the detection rules: https://threadlinqs.com/?modal=signup

## Sources

- [Maintainers don't need more AI bug reports, they need fixes (GitHub Blog)](https://github.blog/open-source/maintainers/maintainers-dont-need-more-ai-bug-reports-they-need-fixes/)
- [Microsoft's AI Uncovers 99 Security Issues in Express, ImageMagick and Other Open-Source Projects (Cyber Security News)](https://cybersecuritynews.com/microsofts-ai-uncovers-99-security-issues/)
- [Defense at AI speed: Microsoft's new multi-model agentic security system (Codename MDASH launch)](https://www.microsoft.com/en-us/security/blog/2026/05/12/defense-at-ai-speed-microsofts-new-multi-model-agentic-security-system-tops-leading-industry-benchmark/)
- [Beyond the benchmark: Advancing security at AI speed (Microsoft Security Blog)](https://microsoft.com/en-us/security/blog/2026/06/17/beyond-the-benchmark-advancing-security-at-ai-speed)
- [Codename MDASH AI code security overview (Microsoft Learn)](https://learn.microsoft.com/en-us/security-exposure-management/ai-code-security-overview)
- [AI-supported vulnerability triage with the GitHub Security Lab Taskflow Agent](https://github.blog/security/ai-supported-vulnerability-triage-with-the-github-security-lab-taskflow-agent/)
- [GitHub Secure Open Source Fund](https://github.com/open-source/github-secure-open-source-fund)
- [Microsoft FORGE Lab blog](https://forge-blog.github.io/)
- [Microsoft unveils MDASH, its AI agent-driven security platform (TechRadar)](https://www.techradar.com/pro/security/microsoft-unveils-mdash-its-ai-agent-driven-security-platform-and-its-already-spotted-a-host-of-new-windows-flaws)

## Full data

The other 8 detection rules (Splunk SPL / Microsoft KQL / Sigma) unlock with TL Creds; a free account includes daily creds: https://threadlinqs.com/?modal=signup
Indicator values unlock with TL Creds, at a lower price on Red and Purple; AI agents can query rules and IOCs through the Threadlinqs MCP server (Purple): https://intel.threadlinqs.com/mcp

Canonical: https://intel.threadlinqs.com/threat/TL-2026-6387
