Skip to content
CVE-2026-35002

Agno FunctionCall field_type eval() Arbitrary Code Execution (CVE-2026-35002)

Execution Last updated:

Detects exploitation of CVE-2026-35002, an arbitrary code execution vulnerability in the Agno AI agent framework (Python, pip package 'agno') prior to version 2.3.24. The field_type parameter on a FunctionCall input field was passed to Python's eval() (CWE-95), allowing an attacker who can influence that value to execute arbitrary Python code in the Agno process. Exploitation typically appears as a Python interpreter spawning unexpected child processes (shells, curl/wget, reverse-shell payloads) or performing unusual network or file activity. This detection surfaces vulnerable agno installations and the process and behavioral signatures of successful code execution. CVSS 9.8, PoC public.

Based on VulnCheck Advisory: agno field_type eval injection arbitrary code execution

Vulnerability Intelligence

Public PoC

CVSS

9.8
Critical (9.0–10)
Attack Vector
Network
Attack Complexity
Low
Privileges Required
None
User Interaction
None
Scope
Unchanged
Confidentiality
High
Integrity
High
Availability
High
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H

What is CVE-2026-35002 Agno FunctionCall field_type eval() Arbitrary Code Execution (CVE-2026-35002)?

Agno FunctionCall field_type eval() Arbitrary Code Execution (CVE-2026-35002) (CVE-2026-35002) maps to the Execution tactic — the adversary is trying to run malicious code or instructions on a system in MITRE ATT&CK.

This page provides production-ready detection logic for Agno FunctionCall field_type eval() Arbitrary Code Execution (CVE-2026-35002), covering the data sources and telemetry it touches: Microsoft Defender for Endpoint. The queries below are rated high severity at medium confidence, and ship for 7 SIEM platforms — KQL, SPL, Elastic, QRadar, Sumo, YARA-L, LogScale.

MITRE ATT&CK

Tactic
Execution
Microsoft Sentinel / Defender
kusto
// df00tech DF00-6a0f0708c3 · https://df00tech.com/detections/CVE-2026-35002 · licence: https://df00tech.com/detections/licence
// Detect Python agent process (agno) spawning suspicious child processes consistent with eval() injection
let suspiciousChildren = dynamic(["sh","bash","zsh","dash","cmd.exe","powershell.exe","pwsh","curl","wget","nc","ncat","socat","perl","ruby"]);
DeviceProcessEvents
| where InitiatingProcessFileName in~ ("python","python3","python.exe","pythonw.exe")
| where InitiatingProcessCommandLine has "agno" or InitiatingProcessCommandLine has_any ("agent","FunctionCall")
| where FileName in~ (suspiciousChildren)
    or ProcessCommandLine has_any ("/dev/tcp","base64 -d","os.system","subprocess","__import__","socket.socket")
| project Timestamp, DeviceName, AccountName, InitiatingProcessFileName, InitiatingProcessCommandLine, FileName, ProcessCommandLine, FolderPath
| order by Timestamp desc

Flags a Python process associated with agno spawning shells, network tools, or executing inline code-execution primitives indicative of successful eval() injection.

high severity medium confidence

Data Sources

Microsoft Defender for Endpoint

Required Tables

DeviceProcessEvents

False Positives

  • Legitimate agno tools that intentionally shell out (e.g. a ShellTool or system-command tool configured by the operator)
  • Developers running agno agents interactively that invoke subprocesses as part of normal workflows
  • CI/CD pipelines executing agno test suites that spawn child processes

Query fingerprint DF00-6a0f0708c3 · licence

Sigma logsource & community rules

The detection logic for Agno FunctionCall field_type eval() Arbitrary Code Execution (CVE-2026-35002) (CVE-2026-35002) above is provided in a vendor-neutral form so you can deploy it on any SIEM. The same logic is shipped here as native KQL (Microsoft Sentinel / Defender), SPL (Splunk), Elastic (Elastic Security (EQL)), QRadar (IBM QRadar (AQL)), Sumo (Sumo Logic CSE), YARA-L (Google Chronicle / SecOps), LogScale (CrowdStrike LogScale (CQL)) queries. In Sigma terms, this detection targets the following logsource:

logsource:
  category: process_creation
  product: windows

Browse the community-maintained Sigma rules for this technique:


Testing Methodology

Validate this detection against 3 adversary techniques from Atomic Red Team. Each test below lists the behaviour to exercise and the telemetry you should expect to see. Executable commands and cleanup steps are further down this page.

  1. Test 1 Simulate agno field_type eval() injection spawning a shell

    Expected signal: Process-creation event: python3 (argv agno-agent) as parent spawning sh -c 'id' with output redirection.

  2. Test 2 Simulate eval() injection performing network egress

    Expected signal: python (agno) parent spawning curl to a localhost URL.

  3. Test 3 Identify vulnerable agno installation

    Expected signal: pip/python process reading agno package metadata.


Response Playbook

Triage

  1. Confirm the affected host runs the agno Python package and determine the installed version with `pip show agno` (or inspect the project's lockfile/requirements). Versions < 2.3.24 are vulnerable.
  2. Review the alerting python process's parent/child chain: identify whether an agno agent spawned the suspicious child process and capture the full command line of both.
  3. Inspect the agent's tool/function schema definitions and any recent inputs (prompts, tool outputs, external schemas) for attacker-controlled field_type values containing Python expressions.
  4. Determine whether the spawned child process performed network egress, file writes, or persistence actions that indicate successful code execution rather than a benign tool invocation.

Containment

  1. Isolate the affected host from the network to prevent further command-and-control or lateral movement.
  2. Terminate the compromised agno agent process and any child processes it spawned, then disable the agent service until patched.
  3. Upgrade agno to version 2.3.24 or later across all affected environments and restrict which tool/function schemas the agent will accept from untrusted sources.

Evidence Collection

  1. Capture the full process tree (parent python/agno process and all children) with command lines, timestamps, and user context.
  2. Preserve the agno agent configuration, tool/function definitions, and any session transcripts or logged FunctionCall objects containing the malicious field_type.
  3. Collect host network connection logs and any files created or modified by the spawned processes for forensic analysis.

Escalation Criteria

  • ! Escalate to incident response if the spawned process established outbound C2, downloaded additional payloads, or achieved persistence.
  • ! Escalate if evidence shows lateral movement, credential access, or data exfiltration originating from the compromised agent host.
  • ! Escalate if multiple hosts or production AI-agent services are confirmed to run vulnerable agno versions and are internet-exposed.

Investigation Guide

Forensic Artifacts

  • > Process-creation events showing a python/agno parent spawning shells or network utilities
  • > agno agent tool/function schema definitions and logged FunctionCall objects containing a malicious field_type
  • > Installed agno package version metadata (pip, site-packages dist-info)

Tuning Guidance

Baseline the legitimate tools your agno agents are configured to use. If an agent is intentionally granted a shell or command-execution tool, exclude that specific child-process/command pattern and instead alert on deviations (unexpected binaries, /dev/tcp reverse shells, base64-decoded commands). Scope the parent-process match to hosts known to run agno to reduce noise, and raise confidence to high once agno version < 2.3.24 is confirmed on the host.


Hunting Queries

Hunt for agno-associated python processes spawning shells or network tools across the estate.

Hunting — KQL
kql
// df00tech DF00-6a0f0708c3 · https://df00tech.com/detections/CVE-2026-35002 · licence: https://df00tech.com/detections/licence
DeviceProcessEvents | where InitiatingProcessFileName in~ ("python","python3") | where InitiatingProcessCommandLine has "agno" | where FileName in~ ("sh","bash","curl","wget","nc") | project Timestamp, DeviceName, InitiatingProcessCommandLine, FileName, ProcessCommandLine
Hunting — SPL
spl
``` df00tech DF00-6a0f0708c3 · https://df00tech.com/detections/CVE-2026-35002 · licence: https://df00tech.com/detections/licence ```
index=* sourcetype="linux:process" parent_process="*python*" parent_process="*agno*" | search process_name IN ("sh","bash","curl","wget","nc") | table _time host parent_process process_name process

Atomic Red Team Tests

Test 1 Simulate agno field_type eval() injection spawning a shell
linux

In a lab, run a python process named/argv'd to resemble an agno agent that evaluates attacker-controlled field_type, causing a shell command to execute.

Command

bash
python3 -c "import sys; sys.argv[0]='agno-agent'; field_type=\"__import__('os').system('id > /tmp/agno_poc.txt')\"; eval(field_type)"

Cleanup

bash
rm -f /tmp/agno_poc.txt

Expected Telemetry

Process-creation event: python3 (argv agno-agent) as parent spawning sh -c 'id' with output redirection.

Expected Detection

KQL/SPL/EDR rules match the python/agno parent spawning a shell via os.system/__import__.

Test 2 Simulate eval() injection performing network egress
linux

Lab simulation where the injected field_type triggers an outbound curl, mimicking C2 download after code execution.

Command

bash
python3 -c "import os; os.environ['AGENT']='agno'; field_type=\"__import__('subprocess').call(['curl','-s','http://127.0.0.1:8000/poc'])\"; eval(field_type)"

Cleanup

bash
echo 'no artifacts to clean'

Expected Telemetry

python (agno) parent spawning curl to a localhost URL.

Expected Detection

Rules match the curl/network-tool child of a python/agno parent.

Test 3 Identify vulnerable agno installation
linux

Enumerate the installed agno version to confirm exposure to CVE-2026-35002 (< 2.3.24).

Command

bash
pip show agno 2>/dev/null | grep -i version || python3 -c "import agno,sys; print(getattr(agno,'__version__','unknown'))"

Cleanup

bash
echo 'no changes made'

Expected Telemetry

pip/python process reading agno package metadata.

Expected Detection

Vulnerability-management correlation flags agno < 2.3.24 as affected (not an EDR behavioral detection).

Related Detections

Blog posts about this detection

Browse by tactic

Full ATT&CK matrix Search all detections