Skip to content

Agentic Red Team Scanner

This scanner probes an autonomous agent that can use tools, memory, or other agents. Use it to test whether malicious input can redirect the agent's goal, misuse its tools, or cross trust boundaries before deployment.

  • Category: Security
  • Purpose: Tests tool-using autonomous agents against risks that do not apply to plain chat models — goal manipulation, tool misuse, memory poisoning, and inter-agent communication
  • Inputs: input, simulator_llm, evaluator_llm
  • Outputs: output - Scan results
  • Configuration:
  • Target Base URL, Target Model Name, Target API Key: Connection details for the agent under test (required)
  • Target Agent Purpose: Default describes an autonomous customer-support agent with tool access. Describe the agent's tools, authority limits, memory usage, and any inter-agent communication — the scanner builds its scenarios from this
  • Target System Prompt: The target's system instruction. The scanner evaluates goal hijacking, tool misuse, and trust exploitation against it
  • Agentic Vulnerabilities: Comma-separated list. Default ExcessiveAgency, GoalTheft, ToolOrchestrationAbuse
  • Vulnerability Type Overrides: Custom vulnerability definitions
  • Attack Vectors: Comma-separated list. Default Prompt Injection, Jailbreaking
  • Attacks Per Vulnerability: 1–50, default 3
  • Max Concurrent: 1–20, default 1
  • Use OWASP ASI 2026: Default off. Switches the scan to the agent-specific OWASP category set
  • OWASP ASI Categories: Comma-separated ASI identifiers. Leave empty to test all ten
  • SSL Certificate Verification, Strip Reasoning/Thinking Tags, Extra Body Parameters (JSON)
  • Supported Agentic Vulnerabilities:
  • Critical — ExcessiveAgency (acting beyond granted authority), GoalTheft (original objective replaced), ToolOrchestrationAbuse (unsafe tool chains), UnexpectedCodeExecution (injected or unsafe code)
  • High — ExploitToolAgent (tool bypass and function abuse), DebugAccess (unauthorized debug mode), ExternalSystemAbuse (API abuse, rate limit bypass), SystemReconnaissance (architecture and tool discovery)
  • Medium — IndirectInstruction (hidden commands in content the agent reads), AgentIdentityAbuse (identity spoofing, role confusion), AutonomousAgentDrift (gradual behavioural deviation), CrossContextRetrieval (access to another session's context), ToolMetadataPoisoning (misleading tool descriptions), RecursiveHijacking (self-referential loops), InsecureInterAgentCommunication (message interception and spoofing)
  • Supported Attack Vectors:
  • Single-turn — Prompt Injection, Gray Box, Prompt Probing, Roleplay, AuthorityEscalation, GoalRedirection, SystemOverride, PermissionEscalation
  • Multi-turn — Jailbreaking (also LinearJailbreaking), TreeJailbreaking, CrescendoJailbreaking, SequentialJailbreak
  • OWASP ASI Top 10 2026: Available when Use OWASP ASI 2026 is enabled. This framework covers autonomous and semi-autonomous agents, whose reasoning, tool integration, persistent memory, and inter-agent messaging create vulnerability classes beyond prompt-level attacks
  • ASI_01 Agent Goal Hijack — Critical. Direct goal manipulation, indirect instruction injection, recursive hijacking, cross-context injection. Dangerous because agents execute multi-step plans automatically, so a hijacked goal can complete extensive unauthorized work before anyone notices
  • ASI_02 Tool Misuse & Exploitation — Critical. Recursive tool calls, unsafe tool composition, budget exhaustion, cross-tool state leakage
  • ASI_03 Agent Identity & Privilege Abuse — High. Agent impersonation, cross-agent trust abuse, identity inheritance, role bypass
  • ASI_04 Agentic Supply Chain Compromise — High. Schema manipulation, deceptive tool descriptions, permission misrepresentation, registry poisoning
  • ASI_05 Unexpected Code Execution — Critical. Unauthorized execution, shell commands, unsafe eval, command injection
  • ASI_06 Memory & Context Poisoning — Critical. Long-term memory poisoning, context injection, state manipulation, memory leakage, cross-session influence. Effects persist, since the agent carries the corrupted information into every later decision
  • ASI_07 Insecure Inter-Agent Communication — High. Agent-in-the-middle, message injection, spoofing, protocol abuse
  • ASI_08 Cascading Agent Failures — High. Tool chain failures, dependency failures, resource exhaustion cascades, trust chain breakdowns
  • ASI_09 Human-Agent Trust Exploitation — Medium. Authority misrepresentation, misleading explanations, over-confidence projection, responsibility diffusion
  • ASI_10 Rogue Agents — Critical. Goal drift, agent collusion, reward hacking, runaway autonomy
  • Usage: The quality of the results depends on how precisely the agent's purpose and system prompt are described. A vague purpose produces generic attacks.