Skip to navigation

Threat reportThreat IntelligenceTL-2026-6446

Prompt Injection (OWASP LLM01) against AI agents: techniques, real-world cases incl. CVE-2026-26030, CVE-2025-32711, GuardBreaker/UAC-0099

highACTIVE

Prompt injection (OWASP LLM01, MITRE ATLAS AML.T0051) lets untrusted content steer LLM agents into leaking data, executing code or compromising software supply chains. Sourced cases include Semantic Kernel RCE (CVE-2026-26030), EchoLeak (CVE-2025-32711), the Gemini CLI and OpenAI Codex CI/CD chains, and in-the-wild use by UAC-0099 of a GuardBreaker comment in a VBS downloader for MATCHBOIL to derail AI-assisted malware analysis. First disclosed .

Key facts

IdentifiersTL-2026-6446, CVE-2026-26030, CVE-2025-32711, CVE-2026-25592
Also known asOWASP LLM01, LLM01 Prompt Injection, GuardBreaker, EchoLeak, Agent Goal Hijack (ASI01)
SeverityHIGH
CVSS9.9 (CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H)
First disclosed
EPSSCVE-2025-32711: 8.0% (94.5th percentile); CVE-2026-26030: 3.7% (89.7th percentile); CVE-2026-25592: 2.3% (83.2th percentile)
Affected productsMicrosoft Semantic Kernel (Python SDK); Microsoft Semantic Kernel (.NET SDK, SessionsPythonPlugin); Microsoft Microsoft 365 Copilot; OpenAI openai/codex GitHub workflows (Issue Labeler, codex-action)
Weaknesses (CWE)CWE-94
AttributionUAC-0099 (medium confidence)
Nation-state nexusRussia
MalwareMATCHBOIL
ToolsCacheract
MITRE ATT&CK / ATLAS techniquesAML.T0051 LLM Prompt Injection, AML.T0051.000 LLM Prompt Injection: Direct, AML.T0051.001 LLM Prompt Injection: Indirect, T1027 Obfuscated Files or Information, T1059.004 Command and Scripting Interpreter: Unix Shell, T1059.005 Command and Scripting Interpreter: Visual Basic, T1059.006 Command and Scripting Interpreter: Python, T1140 Deobfuscate/Decode Files or Information, T1190 Exploit Public-Facing Application, T1195 Supply Chain Compromise, T1203 Exploitation for Client Execution, T1213 Data from Information Repositories and 4 more in the attack flow
MotivationESPIONAGE
Target sectorstechnology, software-development, government administration, energy, transport
Target regionsukraine, Global
Detection rules9 (1 shown free)
Sources16 cited sources

Detection coverage

As of 2026-11-19, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-6446 across Splunk SPL, Microsoft KQL and Sigma, covering 11 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.

Free detection: Python Agent Host Spawning Shell or calc.exe — Semantic Kernel CVE-2026-26030 Exploitation

Detects a Python host process (python.exe/pythonw.exe) spawning calc.exe, a shell or a LOLBin. CVE-2026-26030 (Semantic Kernel Python < 1.39.4, CVSS 9.9) lets a model-controlled filter reach eval() in the InMemoryVectorStore; Microsoft's PoC launched calc.exe from the agent host. Any prompt-injection-to-RCE chain on a Python agent surfaces as this parent/child pair.

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • ATT&CK: T1203, T1059.006, T1190
  • Data sources: sysmon, windows_security, defender_for_endpoint
  • Severity: critical
  • Confidence: high

Splunk SPL

index=windows sourcetype="XmlWinEventLog:Microsoft-Windows-Sysmon/Operational" EventCode=1
(ParentImage="*\\python.exe" OR ParentImage="*\\pythonw.exe")
(Image="*\\calc.exe" OR Image="*\\cmd.exe" OR Image="*\\powershell.exe" OR Image="*\\pwsh.exe" OR Image="*\\wscript.exe" OR Image="*\\cscript.exe" OR Image="*\\mshta.exe" OR Image="*\\certutil.exe" OR Image="*\\rundll32.exe")
```comment: Python process spawning a shell or calc.exe matches the CVE-2026-26030 proof of concept```
| stats count min(_time) as first_seen max(_time) as last_seen values(CommandLine) as child_commands by host, User, ParentImage, ParentCommandLine, Image
| eval first_seen=strftime(first_seen,"%Y-%m-%dT%H:%M:%S"), last_seen=strftime(last_seen,"%Y-%m-%dT%H:%M:%S")
| eval threat_ref="TL-2026-6446 CVE-2026-26030 Semantic Kernel InMemoryVectorStore eval() RCE"

Microsoft KQL

DeviceProcessEvents
| where Timestamp > ago(24h)
| where InitiatingProcessFileName in~ ("python.exe", "pythonw.exe")
| where FileName in~ ("calc.exe", "cmd.exe", "powershell.exe", "pwsh.exe", "wscript.exe", "cscript.exe", "mshta.exe", "certutil.exe", "rundll32.exe")
| project Timestamp, DeviceName, AccountName, InitiatingProcessFileName, InitiatingProcessCommandLine, FileName, ProcessCommandLine
| order by Timestamp desc

Sigma

title: Python Process Spawning Shell or calc.exe - Semantic Kernel CVE-2026-26030
id: 3b7a1c52-8d4e-4f06-9a21-5c0e7d9b1a01
status: experimental
description: Detects python.exe spawning calc.exe, a command interpreter or a LOLBin, the behaviour of the CVE-2026-26030 Semantic Kernel InMemoryVectorStore eval() RCE and other prompt-injection-to-RCE chains on Python agents.
references:
    - https://intel.threadlinqs.com
    - https://nvd.nist.gov/vuln/detail/CVE-2026-26030
    - https://www.microsoft.com/en-us/security/blog/2026/05/07/prompts-become-shells-rce-vulnerabilities-ai-agent-frameworks/
author: Threadlinqs Team
date: 2026-11-19
tags:
    - attack.execution
    - attack.t1203
    - attack.t1059.006
    - attack.initial-access
    - attack.t1190
    - cve.2026-26030
logsource:
    category: process_creation
    product: windows
detection:
    selection_parent:
        ParentImage|endswith:
            - '\python.exe'
            - '\pythonw.exe'
    selection_child:
        Image|endswith:
            - '\calc.exe'
            - '\cmd.exe'
            - '\powershell.exe'
            - '\pwsh.exe'
            - '\wscript.exe'
            - '\cscript.exe'
            - '\mshta.exe'
            - '\certutil.exe'
            - '\rundll32.exe'
    condition: selection_parent and selection_child
falsepositives:
    - Python-based administration or build scripts that legitimately launch shells
    - Developer test runs
level: high

Known false positives:

  • Python-based admin or build scripts that legitimately launch cmd.exe or powershell.exe
  • Developers running Python test suites that open calc.exe or shells

Tuning notes: Upgrade semantic-kernel Python to 1.39.4 or later. Exclude known build/automation parent command lines and service accounts; raise priority when ParentCommandLine contains semantic_kernel, uvicorn, gunicorn or an agent framework. Requires Sysmon EventCode 1 (or MDE) on agent hosts.

8 more detection rules (Splunk SPL, Microsoft KQL, Sigma) · unlock with TL Creds. Sign up free to unlock the detection rules or query them with the Threadlinqs MCP server (Purple)
  • Base64+Gzip Encoded bash -lc Payload on CI Runner — Codex Issue Injection Guardrail Evasion · unlock with TL Creds
  • Cacheract and test_depi.py Cache-Poisoning Artifacts — Codex CI Supply Chain Chain · unlock with TL Creds
  • Python Agent Host Spawning Shell or Download Utility on Linux — Prompt-Injection RCE and Tool Transfer · unlock with TL Creds
  • Cursor Rules File Written by Script Interpreter — Rules File Backdoor · unlock with TL Creds
  • Gemini CLI Running in --yolo Mode — Unrestricted Agent on Untrusted Content · unlock with TL Creds
  • Unicode Tag Character ASCII Smuggling in URLs or Prompt-Injection Script URL — LLM Exfiltration · unlock with TL Creds
  • CoreCLR Debug Pipe and Runner.Worker Memory Read Artifacts — Codex CI Chain Post-Exploitation · unlock with TL Creds
  • VBScript Downloader Spawning Download or Analysis-Tool Checks — UAC-0099 MATCHBOIL GuardBreaker · unlock with TL Creds

Indicators of compromise

As of 2026-11-19, Threadlinqs Intelligence tracks 11 indicators of compromise for TL-2026-6446.

Sample indicators (1 of 11)

  • URL: hxxps://pillar[.]security/security.js

Defanged, TLP:CLEAR. Full set of 11 indicators: unlock with TL Creds.

11 indicators of compromise · unlock with TL Creds. Compare plans

Technical analysis

Prompt injection exploits the fact that LLMs read instructions and data as one token stream, so content from any untrusted channel (email, web page, document, issue, rules file, tool output, image) can be interpreted as instructions. The Picus Security overview (2026-11-19) classifies attacks by delivery surface (direct, indirect, multimodal, clipboard), encoding (authority spoofing, Base64/hex, runtime assembly, Unicode Tag/RTL obfuscation, payload splitting, multilingual, syntax-based) and propagation (single-shot, multi-step chain, cross-session persistence, self-replicating).

CVE-2026-26030 (Semantic Kernel Python SDK < 1.39.4, NVD CVSS 3.1 9.9, CWE-94): the In-Memory Vector Store default filter builds a Python lambda from model-controlled input and runs it with eval(); the AST blocklist guarding it was bypassable (it missed names such as __name__ and system and did not inspect ast.Subscript nodes). Microsoft's Defender research team (blog dated 2026-05-07) demonstrated a hotel-finder agent whose search_hotels city parameter reached the filter and launched calc.exe on the host. A companion flaw, CVE-2026-25592 (.NET SDK < 1.71.0), exposed an unvalidated DownloadFileAsync file-write as a model-callable KernelFunction. The fix adds an AST node-type allowlist, a function-call allowlist and a dangerous-attribute blocklist.

CVE-2025-32711 (EchoLeak, Microsoft 365 Copilot): a zero-click indirect injection delivered through an external email; Cato Networks' write-up describes RAG spraying (repeating the instruction under many HR-style topics) to get the email retrieved, then coercion of Copilot into building an outgoing URL carrying the most sensitive data. Microsoft fixed it server-side; sources report no confirmed customer data theft. Other documented exfiltration patterns: Markdown image URLs (Google AI Studio, Copilot Chat, NotebookLM, GPT-4V), links carrying secrets (Slack AI), ASCII smuggling with Unicode Tag characters U+E0000-U+E007F (Copilot 'Copirate', Grok), and hex-encoded instructions (GitLab Duo).

Agentic and supply-chain cases: Invariant Labs showed a public GitHub issue making an agent pull private-repo data into a public PR via the GitHub MCP server (2025-05-26); Pillar Security showed the Rules File Backdoor (invisible Unicode in a Cursor rules file) and a Gemini CLI triage agent running in --yolo mode that read a token from disk and exfiltrated it, enabling a write-scoped workflow (2026-05-05). Depi Security (2026-11-11) described an openai/codex Issue Labeler chain: attacker issue text mimicking the workflow prompt (allow-users '*') led the agent to run shell commands, escape the workspace-write sandbox via CoreCLR debug pipes in /tmp to read Runner.Worker memory, steal runner/GitHub tokens, poison the rust-ci.yml cache (Cacheract) and set a release trap intended to capture an OIDC token for npm trusted publishing of @openai/codex. The authors do not claim a malicious package was published; OpenAI deployed a fix.

In the wild: ESET (WeLiveSecurity, 2026-09-10) reported that Russia-aligned UAC-0099 placed a GuardBreaker comment, a decoy request for guidance on building a nuclear weapon, in a VBS script used against a target in Ukraine. The comment has no runtime effect; it aims to trip an LLM scanner's safety guardrails so the rest of the file is never analyzed. The script downloads and installs MATCHBOIL, a loader used only by UAC-0099 to deliver further payloads. Secondary reporting notes ESET described intent rather than a confirmed evasion success, and that the group also checked for IDA and Wireshark. Picus also cites Google (2026-04-23) reporting a 32% relative rise in malicious web injections between Nov 2025 and Feb 2026. Defensive guidance: treat all channels as untrusted, decode and normalize input (flag zero-width and Tag characters), least-privilege tools with human approval bound to exact arguments, sanitize output and block arbitrary image/link domains, and cross-check AI malware verdicts with static, behavioral and human analysis. Severity reflects mostly patched or research-demonstrated cases plus one limited in-the-wild case; no CVSS is stated for CVE-2025-32711 in the Picus source.

MITRE ATT&CK / ATLAS techniques by tactic:

Execution

AML.T0051 LLM Prompt Injection; AML.T0051.000 LLM Prompt Injection: Direct; AML.T0051.001 LLM Prompt Injection: Indirect; T1059.004 Command and Scripting Interpreter: Unix Shell; T1059.005 Command and Scripting Interpreter: Visual Basic; T1059.006 Command and Scripting Interpreter: Python; T1203 Exploitation for Client Execution

Defense Evasion

T1027 Obfuscated Files or Information; T1140 Deobfuscate/Decode Files or Information; T1497 Virtualization/Sandbox Evasion

Initial Access

T1190 Exploit Public-Facing Application; T1195 Supply Chain Compromise

Collection

T1213 Data from Information Repositories

Credential Access

T1552.001 Unsecured Credentials: Credentials In Files

Exfiltration

T1567 Exfiltration Over Web Service

Privilege Escalation

T1611 Escape to Host

Remediation

Patches

  • semantic-kernel (Python) 1.39.4 for CVE-2026-26030
  • semantic-kernel (.NET) 1.71.0 for CVE-2026-25592
  • Microsoft 365 Copilot CVE-2025-32711 fixed server-side by Microsoft

Immediate actions

  • Upgrade Semantic Kernel Python to 1.39.4 or later and the .NET SDK to 1.71.0 or later; avoid InMemoryVectorStore in production
  • Restrict agent triggers in CI: do not set allow-users '*' on LLM-driven GitHub Actions workflows and never run coding agents in --yolo / unrestricted mode on untrusted issue content
  • Block or sanitize Markdown images and links to arbitrary domains in LLM output
  • Do not rely on a single AI verdict for malware triage; cross-check with static, behavioral and human analysis

Workarounds

  • Avoid the Search Plugin backed by the In-Memory Vector Store in default configuration until upgraded
  • Disable or gate model-callable file download/upload functions that accept model-supplied paths

Longer-term hardening

  • Treat email, web, documents, issues, rules files and tool output as untrusted input; decode Base64/hex, normalize Unicode and flag zero-width and Tag characters (U+E0000-U+E007F)
  • Apply least-privilege tool scopes, read-only data accounts, per-user authorization and human approval bound to exact arguments
  • Compare extracted document text with rendered text to detect hidden content
  • Run recurring adversarial prompt-injection testing mapped to OWASP LLM Top 10, MITRE ATLAS and NIST AI RMF
  • Pin and review GitHub Actions, scope GITHUB_TOKEN/OIDC permissions minimally and verify cache contents before restore

Affected and fixed versions

  • Microsoft — Semantic Kernel (Python SDK)
    Vulnerable versions: < 1.39.4
    Fixed in: 1.39.4
  • Microsoft — Semantic Kernel (.NET SDK, SessionsPythonPlugin)
    Vulnerable versions: < 1.71.0
    Fixed in: 1.71.0
  • Microsoft — Microsoft 365 Copilot
    Vulnerable versions: pre-fix service (EchoLeak)
    Fixed in: server-side fix by Microsoft
  • OpenAI — openai/codex GitHub workflows (Issue Labeler, codex-action)
    Vulnerable versions: configuration at time of Depi report
    Fixed in: fix deployed by OpenAI

Timeline

  • Greshake et al. publish arXiv:2302.12173 demonstrating indirect prompt injection, cross-session persistence and self-replicating payloads against LLM-integrated applications.
  • PromptArmor shows a public-channel message can steer Slack AI to place a private API key inside a link.
  • Pillar Security discloses the Rules File Backdoor using invisible Unicode in a Cursor rules file.
  • Invariant Labs shows a public GitHub issue can make an MCP-connected agent leak private repository data into a public pull request.
  • Cato Networks publishes its breakdown of EchoLeak (CVE-2025-32711), a zero-click prompt injection against Microsoft 365 Copilot via external email.
  • NVD publishes CVE-2026-26030 (Semantic Kernel Python InMemoryVectorStore filter RCE, CVSS 9.9), fixed in python-1.39.4.
  • Pillar Security details a Gemini CLI issue-triage agent chain (--yolo mode, token read from disk) leading to supply-chain compromise.
  • Microsoft publishes 'When prompts become shells', demonstrating prompt injection to host RCE (calc.exe) in Semantic Kernel (CVE-2026-26030, CVE-2026-25592).
  • ESET reports UAC-0099 embedding a GuardBreaker nuclear-weapon decoy comment in a VBS downloader for MATCHBOIL used against a target in Ukraine.
  • Depi Security publishes the openai/codex Issue Labeler prompt-injection chain toward a potential malicious @openai/codex npm release; OpenAI deployed a fix.
  • Picus Security publishes its OWASP LLM01 overview consolidating these cases and citing a 32% relative rise in malicious web injections between Nov 2025 and Feb 2026.

Sources cited

Showing 15 of 16 sources, primary sources first.

More from UAC-0099

Other reports on CVE-2026-26030

Other reports on CVE-2025-32711

Further reading

Drafted by the Threadlinqs research pipeline from 16 sources; published ; last updated —.

Threadlinqs Intelligence — Real-Time Threat Detection Platform

// threat_feed
$ sort --newest
Showing all threats

Live intelligence console

Threat level

Latest Threats