Introduction
Google has taken immediate action by deleting three AI agent workflows from its Agent Development Kit (ADK) Python repository after security researchers at Pillar Security disclosed a critical prompt injection vulnerability. The flaw allows malicious actors to manipulate a public GitHub issue into triggering a privileged code-fixing agent with escalated permissions. This represents a new attack vector in the AI supply chain that organizations must address immediately. If you're using Google ADK workflows or integrating AI agents with your development repositories, your environment may be at risk of unauthorized code modifications and privilege escalation through seemingly benign user interactions.
Technical Analysis
Affected Products and Components:
- Google Agent Development Kit (ADK) Python repository
- Three specific AI agent workflows (now deleted from the repository)
- GitHub integration components utilizing ADK agents
Vulnerability Mechanics:
The vulnerability stems from an insecure trust relationship between a public triage agent and a privileged code-fixing agent. Researchers demonstrated that a malicious GitHub issue could prompt-inject the triage agent into posting the string /adk-issue-fix as the adk-bot account. Since the bot was configured as a collaborator with elevated permissions, this automated comment satisfied the condition required to trigger the privileged code-fixing agent workflow.
This represents a classic cross-privilege boundary attack in an AI/LLM context:
- Attacker creates a seemingly normal GitHub issue
- The issue contains crafted text that manipulates the triage agent's behavior
- Through prompt injection, the triage agent is coerced into posting a specific command (
/adk-issue-fix) - The privileged agent recognizes this command from a trusted collaborator (itself) and executes code-fixing operations with elevated permissions
Exploitation Status:
- Proof of Concept: Demonstrated by Pillar Security researchers
- Active Exploitation: Not confirmed in the wild, but deletion of workflows suggests urgent remediation
- Exploit Complexity: Low - requires only GitHub issue creation capability
- Privileges Required: None - any GitHub user can create an issue
Detection & Response
SIGMA Rules
---
title: Potential ADK Prompt Injection via GitHub Comment
id: 550d9a23-0f72-4e3a-a8c5-1d2f3b4c5d6e
status: experimental
description: Detects GitHub API activity containing suspicious ADK-related command strings that may indicate prompt injection attempts targeting AI agent workflows.
references:
- https://github.com/google/agent-development-kit
author: Security Arsenal
date: 2026/08/15
tags:
- attack.initial_access
- attack.t1190
logsource:
category: web
product: github
detection:
selection:
cs-method|contains: 'POST'
cs-uri-stem|contains: '/issues/comments'
sc-content|contains:
- '/adk-issue-fix'
- 'adk-bot'
condition: selection
falsepositives:
- Legitimate ADK bot activity
- Authorized testing workflows
level: high
---
title: Suspicious ADK Python Workflow Execution
id: 7a3f1c82-9e4b-4d67-bc12-3e5a8f901234
status: experimental
description: Detects execution of ADK Python workflows with GitHub-related parameters that may indicate exploitation of vulnerable agent workflows.
references:
- https://github.com/google/agent-development-kit
author: Security Arsenal
date: 2026/08/15
tags:
- attack.execution
- attack.t1059.006
logsource:
category: process_creation
product: linux
detection:
selection:
Image|endswith: '/python3'
CommandLine|contains:
- 'adk'
- 'github'
CommandLine|contains:
- 'issue'
- 'comment'
- 'fix'
condition: selection
falsepositives:
- Legitimate ADK development workflows
- Authorized agent operations
level: medium
---
title: GitHub Token Usage by ADK Workflows
id: 9b4e2d93-0f5c-4e78-bc23-4f6b9g012345
status: experimental
description: Detects authentication events using GitHub tokens in conjunction with ADK workflow execution, potentially indicating exploited agent privilege escalation.
references:
- https://github.com/google/agent-development-kit
author: Security Arsenal
date: 2026/08/15
tags:
- attack.credential_access
- attack.t1552.001
logsource:
category: process_creation
product: linux
detection:
selection:
Image|endswith: '/python3'
CommandLine|contains:
- 'GITHUB_TOKEN'
- 'GH_TOKEN'
CommandLine|contains:
- 'adk'
condition: selection
falsepositives:
- Legitimate CI/CD workflows
- Authorized ADK operations
level: medium
KQL for Microsoft Sentinel/Defender
// Hunt for GitHub API calls potentially related to ADK prompt injection
let suspiciousCommands = dynamic(["/adk-issue-fix", "adk-bot"]);
GitHubAuditLogs
| where OperationName in ("issue_comment", "create")
| where Action =~ "created"
| project TimeGenerated, Actor, Repository, Action, Details
| where tostring(Details) has_any (suspiciousCommands)
| extend AccountType = case(Actor contains "[bot]", "Bot", "User")
| order by TimeGenerated desc
// Monitor for suspicious Python processes executing ADK-related workflows
DeviceProcessEvents
| where Timestamp > ago(1d)
| where FileName =~ "python3" or FileName =~ "python"
| where ProcessCommandLine has "adk"
and (ProcessCommandLine has "github" or ProcessCommandLine has "issue")
| project Timestamp, DeviceName, AccountName, ProcessCommandLine, InitiatingProcessFileName
| order by Timestamp desc
Velociraptor VQL
-- Hunt for ADK-related Python processes and script executions
SELECT Pid, Name, CommandLine, Exe, Username, CreateTime
FROM pslist()
WHERE Name =~ "python3"
AND CommandLine =~ "adk"
AND (CommandLine =~ "github" OR CommandLine =~ "issue")
-- Scan for ADK workflow files that may contain vulnerable configurations
SELECT FullPath, Size, Mtime, Data
FROM glob(globs="/*agent*development*kit*/**/*.py")
WHERE Data =~ "issue-fix" OR Data =~ "triage.*agent"
Remediation Script (Bash)
#!/bin/bash
# Google ADK Vulnerability Remediation Script
# This script checks for vulnerable ADK workflows and applies remediation
set -e
echo "[+] Checking for Google ADK installations..."
# Find ADK installations
ADK_PATHS=$(find /home /opt /usr/local -type d -name "*agent*development*kit*" 2>/dev/null || true)
if [ -z "$ADK_PATHS" ]; then
echo "[-] No Google ADK installations found."
exit 0
fi
echo "[+] Found ADK installations:"
echo "$ADK_PATHS"
# Check for vulnerable workflow files
echo "[+] Checking for vulnerable workflow files..."
VULN_FILES=$(find $ADK_PATHS -type f -name "*.py" -exec grep -l "issue-fix" {} \; 2>/dev/null || true)
if [ -n "$VULN_FILES" ]; then
echo "[!] Found potentially vulnerable workflow files:"
echo "$VULN_FILES"
# Backup vulnerable files
echo "[+] Creating backups of vulnerable files..."
BACKUP_DIR="/var/backups/adk_remediation_$(date +%Y%m%d_%H%M%S)"
mkdir -p "$BACKUP_DIR"
cp $VULN_FILES "$BACKUP_DIR/" 2>/dev/null || true
# Remove vulnerable workflows
echo "[!] Removing vulnerable workflow files..."
for file in $VULN_FILES; do
echo "Removing: $file"
rm -f "$file"
done
echo "[+] Backups saved to: $BACKUP_DIR"
else
echo "[+] No vulnerable workflow files found."
fi
# Check for ADK bot tokens in environment
echo "[+] Checking for exposed ADK bot tokens..."
if env | grep -i "GITHUB.*TOKEN" | grep -i "adk" > /dev/null; then
echo "[!] WARNING: Exposed ADK-related GitHub tokens found in environment."
echo " Please rotate these tokens immediately."
else
echo "[+] No exposed ADK-related GitHub tokens found."
fi
# Update ADK to latest version if package manager is available
echo "[+] Checking for ADK package updates..."
if command -v pip3 &> /dev/null; then
pip3 list | grep -i "adk" && echo "[+] Attempting to update ADK packages..." && pip3 install --upgrade google-adk 2>/dev/null || echo "[-] No ADK packages found via pip3"
fi
echo "[+] Remediation complete. Please review logs and confirm changes."
echo "[!] IMPORTANT: If you have ADK workflows integrated with GitHub, review and update your agent permissions immediately."
Remediation
Immediate Actions Required:
-
Identify ADK Usage: Inventory all systems and repositories using Google ADK AI agent workflows, particularly those integrated with GitHub repositories.
-
Remove Vulnerable Workflows: Google has already deleted the three vulnerable workflows from the official repository. If you have cloned or forked these workflows:
- Remove any workflow files containing the issue-fix pattern
- Do not reuse these workflow patterns without security review
- Update any local clones to ensure deleted workflows are removed
-
Review Agent Permissions:
- Audit all AI agent accounts (like
adk-bot) and verify they have minimum necessary permissions - Remove collaborator status from agent accounts where not absolutely required
- Implement approval workflows for privileged operations initiated by agents
- Audit all AI agent accounts (like
-
Implement Input Validation:
- Add strict input validation for all text processed by AI agents
- Sanitize GitHub issue content before passing to agents
- Implement allowlists for agent-triggering commands
-
Agent Isolation:
- Separate public-facing agents (like triage agents) from privileged agents
- Require explicit human authorization for cross-privilege operations
- Implement distinct authentication contexts for different agent privilege levels
Configuration Changes:
-
GitHub Integration Hardening:
- Require two-person approval for automated PR/issue operations
- Implement branch protection rules requiring status checks from trusted sources
- Use fine-grained personal access tokens instead of full repository tokens for agents
-
Prompt Injection Defenses:
- Implement prompt templates with explicit instruction separation
- Add delimiter checking to prevent command injection in agent inputs
- Monitor agent outputs for unexpected command patterns
Official Guidance:
- Monitor Google's official Agent Development Kit repository for security updates: https://github.com/google/agent-development-kit
- Review Pillar Security's full disclosure for additional technical details
Verification: After applying remediation:
- Confirm vulnerable workflows are removed from all environments
- Verify agent permissions follow principle of least privilege
- Test that legitimate operations continue to function
- Implement logging for all agent-initiated actions
Long-term Recommendations:
- Establish a security review process before integrating AI agents into CI/CD pipelines
- Implement AI agent monitoring and alerting for unusual behaviors
- Regularly audit agent permissions and access patterns
- Maintain an inventory of all AI agents and their integration points
Related Resources
Security Arsenal Penetration Testing Services AlertMonitor Platform Book a SOC Assessment vulnerability-management Intel Hub
Is your security operations ready?
Get a free SOC assessment or see how AlertMonitor cuts through alert noise with automated triage.