BOSSAI — Taiwan-Based Operator Uses AI Agents to Exfiltrate Taipei Emergency Dispatch Records

Publication date: November 09, 2026
Category: Artificial Intelligence / Cyber Espionage / Automated Offensive Operations

Introduction

Cybersecurity researchers uncovered the operations of a lone threat actor based in Hualien City, Taiwan, who leveraged a customized framework of artificial intelligence agents named BOSSAI to automate reconnaissance and exploitation campaigns against hundreds of organizations across at least 11 countries. The operator successfully accessed and exfiltrated sensitive emergency dispatch records from Taipei’s ambulance and fire services (119), while also compromising corporate and institutional databases through access control flaws, static cryptographic keys, and SQL injection vulnerabilities. The investigation sheds light on a scenario where modified open-source language models (abliterated) were utilized to bypass commercial security guardrails and coordinate large-scale attacks against critical infrastructure sectors.

What is BOSSAI and the AI Agent Ecosystem? (General Analysis)

The framework designated as BOSSAI consisted of an architecture comprising 12 specialized artificial intelligence agents (with 9 active simultaneously), designed to automate the full lifecycle of an offensive operation: from maintaining and provisioning the attacker’s own infrastructure to scanning targets and executing intrusion tactics.

Unlike fully autonomous systems communicating natively across networks, BOSSAI required partial manual intervention during certain phases, where the operator copied and pasted outputs generated by one agent into another. Because commercial AI models (such as Grok and MiniMax M3) rejected explicit attack requests by demanding proof of authorization, the operator heavily relied on local language models stripped of their safety filters (abliterated models), primarily utilizing open-source architectures such as DeepSeek, GLM, and Qwen.

Although no formal CVE identifier is directly assigned to the BOSSAI framework, the techniques exploited by its agents map to well-known weakness classes within the CWE catalog:

  • CWE-798 (Use of Hard-coded Credentials): Reasoned estimation for flaws in administrative panels and exposed AES keys.
  • CWE-89 (SQL Injection): Reasoned estimation for unauthorized access to financial and academic database tables.
  • CWE-522 (Insufficient Credential Management): Reasoned estimation for the use of default root credentials in technology vendor administration panels.

How Does It Work? (Technical Analysis)

The technical operation of the campaign driven by BOSSAI combined automated software engineering workflows with traditional web and infrastructure exploitation techniques:

  • Initial Flow and Infrastructure Development: The operator utilized cloud-based commercial models (primarily Grok) to build and maintain attack servers, auxiliary tools, and exploitation scripts. When commercial models blocked explicit offensive requests, control was diverted to local abliterated models or coding agents such as Claude Code, Codex, and Cursor configured locally.
  • Evasion and Agent Behavior: Agents operated in parallel to perform port scans, audited exposed services, and executed proof-of-concept tests. However, the system exhibited significant operational inaccuracies; scanning tools employed by the agents generated a high rate of false positives, inventing non-existent CVEs (such as a critical Apache remote code execution flaw that never affected the analyzed national disaster response center).
  • Access and Exfiltration Mechanisms: Successful attacks relied not on sophisticated zero-day exploits, but on pre-existing misconfigurations and systemic exposure:
    • Logging into technology vendor database administration panels using default root credentials.
    • Extracting static AES-256 keys hardcoded directly into client-side JavaScript front-end code, compromising session cookies and administrative privileges.
    • Classical SQL injection (SQLi) against vulnerable endpoints across financial and academic institutions in Taiwan and Thailand.
    • Direct exposure of patient data (including ECG readings and 2018 ambulance video footage) stored on platforms lacking mandatory authentication.

“Automated agents met their boundary exclusively in environments where robust authentication, strict session validation, and proper endpoint filtering were enforced. AI automation drastically reduced the cost and time required for reconnaissance, but failed against correctly implemented defense-in-depth controls.”

Affected Systems / Environments

The campaign impacted multiple critical sectors alongside governmental and private institutions, notably:

  • Emergency Services: 119 dispatch systems across Taipei districts (extracting incident times, response types, case status, and street-level locations) and fire department platforms exposing patient medical data.
  • Vendor Infrastructure: Database administration panels belonging to technology vendors servicing emergency operations.
  • Financial and Academic Sectors: Lending firms in Taiwan (compromised via account table SQLi) and multiple universities in Thailand.
  • Demographic Records: Government and telecommunications data in Mexico (exposing home phone records containing names and addresses).
Component Category / Vector Associated CWE (Estimation) Observed Impact CVSS Score (Estimated) CVSS Vector (Estimated)
DB Administration Panel CWE-798 (Hard-coded / Default Credentials) Access to 18+ database tables 8.8 (High) CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H
Web Application (Frontend JS) CWE-798 (Hard-coded Cryptographic Key) Exposure of administrator session cookies 7.5 (High) CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:N/A:N
Database Endpoints CWE-89 (SQL Injection) Exfiltration of account data and academic records 8.6 (High) CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:L/N:L
Medical Data Platform CWE-306 (Missing Authentication) Public exposure of ECGs and ambulance video footage 7.5 (High) CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:N/A:N

Mitigation and Detection

Remediation

  • Removal of Default Credentials: Audit all internet-exposed systems to ensure default administrative accounts (root, admin) are disabled or updated with robust, rotating passwords.
  • Secure Secret Management: Remove cryptographic keys (such as AES or symmetric keys) from client-side JavaScript code and migrate session management to secure server-side token-based mechanisms.
  • Input Validation and Parameterization: Implement parameterized queries and ORMs across all web applications to prevent SQL injection vulnerabilities.
  • Strict Access Control (IAM): Enforce mandatory Multi-Factor Authentication (MFA) across institutional management portals and restrict API access via gateways with Role-Based Access Control (RBAC).

Detection

  • Monitor anomalous HTTP requests targeting public service APIs with automated mass-query patterns.
  • Inspect database access logs for SQL injection patterns (UNION SELECT, anomalous boolean operators).
  • Detect persistent SSH connections or sessions originating from IP addresses associated with intensive automated scanning behaviors.
yaml
title: Automated Agent Reconnaissance Pattern Detection (BOSSAI IOCs)
id: bossai-agent-recon-2026
status: experimental
description: Detects mass-query patterns and port scanning associated with AI-driven automated reconnaissance tools.
logsource:
  category: webserver
  product: nginx
detection:
  selection:
    uri|contains:
      - "/api/v1/dispatch"
      - "/admin/db/login"
    status: 200
  condition: selection
falsepositives:
  - Authorized security audits
level: medium

Wrapping Up

The incident driven by the BOSSAI framework highlights the evolution in capabilities among lone threat actors through the use of modified local artificial intelligence. Although AI agents did not constitute sophisticated zero-day threats, they demonstrated lethal efficiency in automating the discovery and exploitation of trivial vulnerabilities and pre-existing human configuration errors (unprotected public APIs, default credentials, and static keys). The resilience displayed by systems maintaining strict authentication policies underscores that traditional defensive cybersecurity countermeasures remain the most effective barrier against emerging offensive automation.

References

  • Acronis Threat Research Unit. (2026). BOSS-AI: Inside the AI attack mesh that targeted Taiwan’s emergency services. Acronis Technical Analysis. Retrieved from https://www.acronis.com/en/tru/posts/bossai-inside-the-ai-attack-mesh-that-targeted-taiwans-emergency-services/
  • Khandelwal, S. (2026). Taiwan-Based Hacker Used BOSSAI AI Agents to Steal Taipei Emergency Dispatch Records. The Hacker News. Retrieved from https://thehackernews.com/2026/11/taiwan-based-hacker-used-bossai-ai.html