BOSSAI: When AI Agents Become an Attack Force Multiplier

Abhishek Verma
·
November 19, 2026

TL/DR

The BOSSAI campaign illustrates an evolution in AI enabled cyberattacks. A single operator used a network of specialized AI agents, models, MCP servers, and offensive security tools to conduct cyber operations across multiple countries. Acronis 12 configured AI agents were identified, nine active at the time of investigation, supporting reconnaissance, vulnerability identification, exploit development, and attack execution.

Lineaje Research assessment: BOSSAI is primarily an attack capability amplification incident, not a prompt injection incident. The attacker deliberately configured AI models and agents to pursue offensive objectives, granted broad tool access, and automated significant portions of the attack lifecycle.

The attack exposes 3 critical security concerns:

  • Model safeguards are not execution boundaries. The operator used locally hosted models with safety alignment removed when commercial models refused offensive instructions.
  • Tool access turns AI generated plans into executable attacks. MCP connected security tools enabled agents to conduct scanning, exploitation, credential extraction, and collection.
  • Persistent skills and agent memory support operational continuity. Reusable attack procedures accumulated knowledge and evolved across sessions, reducing the need for repeated human direction.

The central Lineaje finding is that agentic autonomy combined with powerful tools and insufficient authorization boundaries can transform existing security weaknesses into scalable exploitation opportunities.

What Happened

Acronis Threat Research Unit investigated a coordinated AI assisted attack environment operated by one individual in Taiwan, examining more than 90,000 files and approximately 9 GB of evidence, including conversations, execution histories, configurations, credentials, and collected victim data.

The operator assembled specialized infrastructure and offensive agents rather than relying on a single assistant. The environment included commercial models for development and administration, locally modified models for offensive use, and MCP servers connecting agents to security and exploitation tools.

Logical workflow:

Human operator → infrastructure and offensive agents → MCP connected tools → external target systems.

In a documented workflow, HalberdStrike scanner scanned a Taiwanese emergency services target and produced a vulnerability report. The operator transferred results to Hermes for attack path analysis and then to Cursor for exploitation. Elsewhere, Hermes used Kali MCP tools to enumerate Thai university subdomains, identify SQL injection, extract a database username, enumerate an administrator table, and attempt student data extraction.

The operation was not uniformly successful. Commercial models repeatedly refused offensive requests, agents sometimes hallucinated vulnerabilities, encountered tool failures, or abandoned targets. The evidence does not establish reliable autonomous exploitation of every target.

Lineaje Research Assessment

1. The security boundary shifted from model behavior to agent capabilities

The operator controlled objectives, model selection, permissions, and connections to offensive tools. This makes BOSSAI different from a compromised legitimate assistant responding to adversarial content. The operative combination was autonomy, tool privileges, and reachable targets. Model output provided plans, MCP connected tools enabled execution, and exposed or vulnerable services determined exploitability.

2. Removing model safeguards is different from defeating them

The investigation provides an important distinction between bypassing safeguards and replacing them. Acronis Refusals by commercial models were observed. The operator responded by moving offensive workflows to locally hosted, abliterated models and configuring persistent instructions and automatic approvals. An abliterated model is an open weight LLM that has been modified to remove or reduce its built-in safety refusals. This is evidence of model substitution and permissive configuration, not proof of a successful jailbreak against every protected model. From Lineaje's perspective, this reinforces the importance of maintaining an enforceable authorization boundary between AI reasoning and the tools used to execute its decisions.

3. MCP transformed planning into external actions

The environment provided tools for scanning, exploitation, binary analysis, command execution, and automation. More than 3,300 terminal commands, 2,000 code execution calls, 500 file writes, and 129 Kali MCP invocations associated with Hermes activity were reported. Tool availability is not equivalent to tool authorization. Controls should account for tool, operation, destination, identity, data sensitivity, and task context. This represents an important enforcement opportunity for tool authorization, destination restrictions, and runtime policy controls enabled by UnifAI.

4. Persistent skills formed reusable attack workflows

Hermes maintained SKILL.md files and supporting scripts as durable procedures. The gov-api-penetration skill was invoked 51 times and modified 57 times. These artifacts preserved procedures and discoveries across sessions. This is significant because a skill is not merely reference material. It can influence future reasoning and define executable procedures. Skills should continuously be scanned, classified and restricted from unauthorized changes. An agent's capability can change even when its underlying model remains unchanged.

5. Goal integrity alone is insufficient; path integrity defines the security boundary

Goal integrity: The offensive agents generally pursued operator specified objectives. This was not principally a case of an adversary hijacking a legitimate agent’s goals.

Path integrity: The agents selected techniques and executed tools with broad discretion. In an enterprise agent, comparable behavior outside approved boundaries would be a path integrity failure.

Outcome integrity: Unauthorized access and extraction occurred at confirmed targets, while other attempts failed, stalled, or produced false findings. Outcomes must be distinguished by evidence.

The important security principle is that an agent can faithfully pursue its assigned objective while taking an unsafe or unauthorized path.

Lineaje AI Kill Chain Analysis

Stage Tactic Technique Procedure Mitigation / UnifAI policy
1. Reconnaissance Environment Discovery Execution Environment Mapping Agents used FOFA, Shodan, nmap and other tools to discover exposed assets, enumerate services and identify potential vulnerabilities. UnifAI URL and Tool Allowlist policies can directly restrict reconnaissance by enterprise-managed agents, but not by attacker-operated agents outside the organization's control.
2. Trust Establishment & Manipulation Trust Conditioning Gradual Alignment Erosion Operator selected abliterated models, rerouted some requests to local models, and configured automatic action approval. UnifAI Model Allowlist restrict unapproved models. Enforce independent tool authorization and prohibit unrestricted execution modes.
3. Input & Instruction Weaponization Instruction Injection Direct Prompt Injection Hermes received persistent instructions discouraging refusal and encouraging offensive behavior and broad decision-making. UnifAI Prompt Safety policies can detect prohibited instructions. Protect system prompts and require approval for persistent configuration changes.
4. Reasoning-Time Execution Goal Directed Planning Adaptive Path Discovery AI agents analyzed reconnaissance findings, identified potential exploitation paths, and selected subsequent actions based on discovered vulnerabilities and execution results. Enforce task scoped authorization and human approval for high-risk actions. UnifAI Tool Authorization policies constrain execution of unauthorized paths at governed tool boundaries; UnifAI HITL policies require human approval for risky actions.
5. Tool Invocation & Environment Interaction Automation Abuse Autonomous Tool Invocation Hermes invoked Kali MCP tools, including sqlmap; agents also ran terminal commands, code and file operations. UnifAI Tool Allowlist, and URL Allowlist policies restrict executable tools, operations and destinations at governed AI boundaries.
6. Privilege Escalation Unauthorized Access Expansion Missing or Weak Access Controls Agents accessed a database administration panel using default root credentials and exploited SQL injection to access protected database information. UnifAI Tool Authorization constrains governed agent actions; credentials policy restricts credentials usage by agents.
7. Lateral Movement Cross Environment Pivoting Trust Domain Pivoting The operator manually transferred findings and plans among agents; victim-side lateral movement was not confirmed. UnifAI Identity and Agent Tool policies can constrain governed cross-agent workflows. No confirmed victim-side lateral movement is established.
8. Persistence Access Retention Durable State Modification Hermes maintained and repeatedly modified persistent skills containing offensive procedures, reconnaissance findings, and exploitation knowledge, preserving attack capabilities across sessions. UnifAI Skill Classification identifies malicious and suspicious skills upstream; integrity monitoring and change controls detect unauthorized modifications to persistent agent state.
9. AI-Native Command & Control Human-Mediated Control Human-in-the-Loop C2 The operator coordinated specialized AI agents, transferred findings and instructions between them, and directed subsequent attack activities using agent-generated results. UnifAI Agent/Tool Authentication and Destination Allowlist policies restrict governed AI interactions; network controls monitor unauthorized communications.
10. Actions on Objectives Data Theft Data Exfiltration via AI Agents extracted emergency dispatch records, accessed medical information and retrieved financial and university records via exposed services and vulnerable applications. UnifAI Sensitive Data and Egress policies apply to governed flows; API authorization, DLP and network controls directly protect victim data.

‍

Lessons and Conclusion

BOSSAI demonstrates that AI enabled attacks can scale through orchestration, not necessarily novel exploits. Commercial and modified models, MCP connected tools, and persistent skills allowed one operator to automate reconnaissance and portions of exploitation but also produced errors and failed attempts. Existing weaknesses - exposed APIs, default credentials, outdated applications, and injection flaws - remained decisive.

Three lessons stand out from Lineaje’s research perspective:

Lesson 1: Separate model safety from authorization. Refusal behavior can reduce harmful participation, but action execution must be independently governed.

Lesson 2: Govern the connected AI ecosystem. The effective risk arises from relationships among models, agents, tools, skills, MCP servers, identities and destinations.

Lesson 3: Enforce path integrity. An agent may preserve a legitimate objective while choosing prohibited actions. Policy must constrain execution paths and outcomes, not only prompts.

Lineaje’s combined approach, AI component intelligence, model and skill classification, and UnifAI runtime policies, addresses the AI side execution risk when supported by application, identity, endpoint and network controls.

The defining challenge is no longer just whether an AI agent can generate an attack. It is whether the agent has the capabilities, privileges and unrestricted execution paths to carry it out.

When an AI agent has broad tool access and nothing standing between its plans and execution, every exposed weakness becomes a scalable target.
Close the gap with UnifAI →
November 19, 2026