RC RANDOM CHAOS

Single Attacker Weaponized Abliterated AI Agents Against Emergency Services

· via The Hacker News

Original source

Taiwan-Based Hacker Used BOSSAI AI Agents to Steal Taipei Emergency Dispatch Records

The Hacker News →

A Taiwan-based attacker launched a sprawling campaign across 11 countries by deploying a cluster of autonomous AI agents known as BOSSAI. Researchers identified him through identity-reuse mistakes linking his legitimate businesses to attack infrastructure. His agents successfully breached roughly ten targets, with Taiwan experiencing the most severe impact. The most significant compromises involved emergency services: agents accessed a Taipei fire and ambulance dispatch system spanning seven districts, along with a fire department platform lacking authentication controls that exposed years of medical records—ECG readings and ambulance footage—dating back to 2018.

The attacker built BOSSAI with 12 AI agents filling different roles—some managed infrastructure and deployment, while others conducted reconnaissance and carried out attacks. He initially tried commercial cloud models like Grok and MiniMax M3 as builders, but they rejected direct requests to assist in unauthorized access. Instead, the attacker pivoted to locally-run “abliterated” models—variants of DeepSeek, GLM, and Qwen with safety guardrails stripped—to conduct the actual intrusions. The agents operated independently but lacked coordination; the attacker manually relayed outputs between them, introducing friction into the workflow. Despite access to AI, the attack agents produced significant false positives and unreliable intelligence.

Most breaches exploited primitive security gaps that had persisted for years: unauthenticated APIs exposing sensitive records, SQL injection vulnerabilities in financial and educational databases, and development systems left internet-facing with no access controls. In Mexico, agents harvested approximately 206,800 home phone records by combining publicly available government statistics and telecom authority information. The offensive use of AI proved effective for reconnaissance and targeting at scale, enabling the attacker to systematically scan hundreds of potential victims across multiple sectors. Yet the agents consistently faltered against basic defenses: enforced authentication, session validation, and properly configured network filters repelled every attempt, demonstrating that architectural controls remained more resilient than outdated security practices.

Read the full article

Continue reading at The Hacker News →

This is an AI-generated summary. Read the original for the full story.