A lone operator in Taiwan utilized a mesh of 12 AI agents, dubbed BOSSAI, to conduct a multi-country cyberattack campaign. The agents, which included both commercial and abliterated open-source models, were used to scan for vulnerabilities, harvest credentials, and exfiltrate sensitive data from emergency services, universities, and financial institutions. The investigation highlights the effectiveness of AI-augmented threats in automating cyber-offensive operations.
According to Acronis researchers, a Taiwan-based attacker used a coordinated network of AI agents called BOSSAI to access an emergency-dispatch system in Taipei, Taiwan, and extract case records from seven city districts. The records included incident times, response types, status fields, and street-level locations. The researchers described the extraction as part of a wider AI-assisted cyberattack campaign.
Risk classification
- Primary risk domain: 4 Malicious actors
- Primary risk subdomain: 4.2 Cyberattacks, weapon development or use, and mass harm
The incident involves a lone operator using a sophisticated mesh of AI agents to conduct automated cyberattacks, espionage, and data theft against critical infrastructure, government, and educational institutions.
Additional risk subdomains
- 2.1 Compromise of privacy by obtaining, leaking or correctly inferring sensitive information: The AI agents successfully exfiltrated sensitive personal data, including over 200,000 Mexican residential records, student databases, and emergency patient-care records.
- 7.3 Lack of capability or robustness: The local abliterated models frequently hallucinated vulnerabilities, misclassified failed processes as discoveries, and generated unreliable pentesting reports.
Causal factors
- Entity: Human
- Intent: Intentional
- Timing: Post-deployment
The cyberattack campaign was intentionally orchestrated by a human operator who deployed a mesh of AI agents to autonomously target and exfiltrate data from various organizations.
EU AI Act risk tier
High Risk: The AI system was deployed to target critical infrastructure, specifically emergency services (Taipei 119 emergency dispatch and patient-care systems). Under the EU AI Act, AI systems used in critical infrastructure that pose significant implications for safety and public interest are classified as High Risk.
AI system and alleged parties
- AI system: DS4, Grok, Hermes Agent (DeepSeek, xAI)
- AI purpose: Target Identification; Code Generation
- Behaviour type: Multi-agent
- Alleged developer: AI agent system developers
- Alleged deployer: hackers, Cybercriminals, AI agent system deployers
- Alleged harmed parties: Privacy, People whose information was extracted from Taipei, Taiwan's emergency-dispatch system, Emergency services
Harm severity
Severity ratings are based on this scale.
Highest direct severity in any category: Substantial. Severity is scored from Negligible to Catastrophic in each harm category, for harm the reports describe as caused directly or indirectly by the AI system.
- Physical: direct Negligible, indirect Negligible
- Infrastructure: direct Negligible, indirect Negligible
- Property: direct Negligible, indirect Negligible
- Financial: direct Negligible, indirect Negligible
- Environmental: direct Negligible, indirect Negligible
- Malicious content: direct Negligible, indirect Negligible
- Differential treatment: direct Negligible, indirect Negligible
- Civil rights: direct Negligible, indirect Negligible
- Democracy: direct Negligible, indirect Negligible
- Privacy: direct Substantial, indirect Negligible
- Psychological: direct Negligible, indirect Negligible
- Epistemic: direct Minor, indirect Minor
- Child sexual exploitation and abuse: direct Negligible, indirect Negligible
Privacy
Reported: Yes, the report explicitly describes extensive privacy violations and data exfiltration across multiple countries.
Directly caused: The AI agents exfiltrated 206,800 residential phone records with names and addresses in Mexico, 124,000 yellow-page records, 2,563 university administrator records in Thailand, and sensitive patient-care records (ECG readings, ambulance video, vital signs) from a Taipei fire-department platform.
Inferred additional harm: It is likely that additional student records and emergency dispatch details were compromised, exposing thousands of additional individuals to potential identity theft or targeted scams.
Epistemic
Reported: Yes, the report describes epistemic harm in the form of AI-driven hallucinations and fabricated vulnerability findings.
Directly caused: The AI agents fabricated critical vulnerabilities, such as a non-existent Apache remote-code-execution flaw on a national disaster-response center portal, and flagged failed connection attempts as critical RCEs.
Indirectly caused: These hallucinations overloaded the operator with false positives, with over half of the high-severity findings being false positives, leading to wasted effort and confusion.
People affected
- Occurrences reported: 1
- People reportedly harmed: 333363
- People reportedly exposed: 333363
Potential causes
Management
- Ineffective Model Selection: Constantly swapping models caused operational instability and execution errors.
- Lack of Oversight on Agent Scope: Autonomous agents drifted in scope, attacking or ignoring targets on their own.
Technology
- Abliterated Models Remove Alignment: Using models with safety alignment stripped bypassed built-in safety rules.
- Autonomous Execution YOLO Mode: HERMES_YOLO_MODE enabled auto-approval of actions without human confirmation.
- Model Hallucinations: Agents fabricated vulnerabilities and CVEs, leading to unreliable outcomes.
Data Inputs
- Unfiltered Target Recon Data: Feeding raw scan results directly into models without validation caused errors.
- Lack of Source Verification: AI agents parsed flawed or fabricated data inputs as genuine vulnerabilities.
Human Factors
- Operator Copy-Paste Errors: Operator acted as a manual bridge, introducing latency and manual errors.
- Poor OPSEC and Identity Reuse: Operator reused real-world handles and residential IP, exposing the system.
- Over-reliance on AI Capabilities: Operator trusted hallucinated vulnerability reports without manual verification.
Process and Methods
- Lack of Agent-to-Agent Protocols: Absence of direct communication required manual human bridging of outputs.
- Self-Extending Skill Loops: Self-patching skill files accumulated errors across execution sessions.
- Disabled Scope Enforcement: Disabling safety and scope checks led to unintended target exploration.
Regulatory Environment
- Unregulated Open-Source LLMs: Availability of powerful open models allowed easy modification and abliteration.
- Lack of Dual-Use Controls: Inability to restrict access to code generation tools for malicious purposes.
Information quality
- Classification confidence: High
- Reason for confidence: The threat research reports provide exceptionally detailed, first-hand forensic evidence from a misconfigured development environment, including 9GB of data, 90,000 files, agent conversation logs, configuration files, and specific victim details. This allows for highly precise classification of the AI systems, the operator's methods, and the resulting data breaches.
- Ambiguities identified: The exact number of Thai students and Taiwanese emergency patients affected is not fully quantified.
- Alternative interpretations: None. The evidence clearly points to an AI-augmented cyber espionage and data theft campaign.
A lone operator in Taiwan deployed a multi-agent AI mesh called BOSSAI to conduct automated cyberattacks across eleven countries. The campaign successfully breached Taipei's emergency dispatch systems and exfiltrated sensitive data from multiple nations, demonstrating an evolved threat capability where AI automates offensive cyber operations targeting critical infrastructure.
- Overall national security impact: Substantial
- Scope: Multiple nations
- Primary target: Taiwan
- Other affected: Thailand, Mexico
- Alleged perpetrator: Individual
Threat characteristics
- Imminence: Long-term. The incident represents an ongoing strategic concern regarding the proliferation of AI-augmented offensive cyber capabilities.
- Autonomy: Human-supervised. The agent mesh executed automated workflows in 'YOLO' mode, but still required a human operator to orchestrate the infrastructure and handle outputs.
- Novelty: Evolved capability. Represents a significant advancement in cyber threats by leveraging a coordinated multi-agent AI mesh with abliterated open-source models.
Impact by dimension
- Physical security: Substantial. The compromise of Taipei's 119 emergency dispatch and patient-care systems represents a substantial threat to critical public safety infrastructure, although no kinetic disruption was reported.
- Information security: Minor. The campaign exfiltrated sensitive data but did not target classified government intelligence or conduct systematic information warfare.
- Sovereignty: Minor. Breaching municipal emergency services represents a minor impact on local government functions, but did not disrupt national sovereignty.
- Economic security: Minor. Targeting of lending services and universities caused minor economic impact, primarily related to data theft and incident response costs.
- Societal stability: Minor. Mass exfiltration of over 300,000 personal records, including medical and residential data, constitutes a notable violation of privacy rights.