Skip to content

DeepTeam is a framework to red team LLMs and AI agents.

.

penguinedb41b151ec04
diff --git a/README.md b/README.md
index cac6ed12..cbcdc252 100644
--- a/README.md
+++ b/README.md
@@ -56,107 +56,125 @@ DeepTeam runs **locally on your machine** and uses LLMs for both attack simulati
# πŸ”₯ Vulnerabilities, Attacks, and Features
- πŸ“ 50+ ready-to-use [vulnerabilities](https://www.trydeepteam.com/docs/red-teaming-vulnerabilities) (all with explanations) powered by **ANY** LLM of your choice. Each vulnerability uses LLM-as-a-Judge metrics that run **locally on your machine** to produce binary pass/fail scores with reasoning:
- <details><summary><b>Data Privacy</b></summary>
-
- - PII Leakage β€” direct disclosure, session leak, database access
- - Prompt Leakage β€” secrets, credentials, system prompt extraction
-
- </details>
- <details><summary><b>Responsible AI</b></summary>
-
- - Bias β€” gender, race, political, religion
- - Toxicity β€” profanity, insults, threats
- - Child Protection β€” CSAM, grooming
- - Ethics β€” moral violations
- - Fairness β€” discriminatory outcomes
-
- </details>
- <details><summary><b>Security</b></summary>
-
- - BFLA β€” broken function-level authorization
- - BOLA β€” broken object-level authorization
- - RBAC β€” role-based access control bypass
- - Debug Access β€” development endpoint exposure
- - Shell Injection β€” command execution attacks
- - SQL Injection β€” database query manipulation
- - SSRF β€” server-side request forgery
- - Tool Metadata Poisoning β€” corrupting tool descriptions
- - Cross-Context Retrieval β€” accessing data across boundaries
- - System Reconnaissance β€” probing agent architecture
-
- </details>
- <details><summary><b>Safety</b></summary>
-
- - Illegal Activity β€” fraud, weapons, drugs
- - Graphic Content β€” violence, sexual content
- - Personal Safety β€” self-harm, dangerous advice
- - Unexpected Code Execution β€” arbitrary code generation
-
- </details>
- <details><summary><b>Business</b></summary>
-
- - Misinformation β€” factual errors, unsupported claims
- - Intellectual Property β€” copyright violations
- - Competition β€” competitor endorsement
-
- </details>
- <details><summary><b>Agentic</b></summary>
-
- - Goal Theft β€” extracting or redirecting an agent's objectives
- - Recursive Hijacking β€” exploiting recursive agent calls
- - Excessive Agency β€” agents acting beyond their authority
- - Robustness β€” input overreliance, hijacking
- - Indirect Instruction β€” hidden instructions in retrieved content
- - Tool Orchestration Abuse β€” exploiting tool calling sequences
- - Agent Identity & Trust Abuse β€” impersonating agent identity
- - Inter-Agent Communication Compromise β€” exploiting multi-agent message passing
- - Autonomous Agent Drift β€” agents deviating from intended goals
- - Exploit Tool Agent β€” leveraging tools for unintended actions
- - External System Abuse β€” using agents to attack external services
-
- </details>
- <details><summary><b>Custom</b></summary>
-
- - Custom Vulnerabilities β€” define and test your own with custom criteria in a few lines of code
-
- </details>
+
+<details>
+<summary><b>Data Privacy</b></summary>
+
+- PII Leakage β€” direct disclosure, session leak, database access
+- Prompt Leakage β€” secrets, credentials, system prompt extraction
+
+</details>
+
+<details>
+<summary><b>Responsible AI</b></summary>
+
+- Bias β€” gender, race, political, religion
+- Toxicity β€” profanity, insults, threats
+- Child Protection β€” CSAM, grooming
+- Ethics β€” moral violations
+- Fairness β€” discriminatory outcomes
+
+</details>
+
+<details>
+<summary><b>Security</b></summary>
+
+- BFLA β€” broken function-level authorization
+- BOLA β€” broken object-level authorization
+- RBAC β€” role-based access control bypass
+- Debug Access β€” development endpoint exposure
+- Shell Injection β€” command execution attacks
+- SQL Injection β€” database query manipulation
+- SSRF β€” server-side request forgery
+- Tool Metadata Poisoning β€” corrupting tool descriptions
+- Cross-Context Retrieval β€” accessing data across boundaries
+- System Reconnaissance β€” probing agent architecture
+
+</details>
+
+<details>
+<summary><b>Safety</b></summary>
+
+- Illegal Activity β€” fraud, weapons, drugs
+- Graphic Content β€” violence, sexual content
+- Personal Safety β€” self-harm, dangerous advice
+- Unexpected Code Execution β€” arbitrary code generation
+
+</details>
+
+<details>
+<summary><b>Business</b></summary>
+
+- Misinformation β€” factual errors, unsupported claims
+- Intellectual Property β€” copyright violations
+- Competition β€” competitor endorsement
+
+</details>
+
+<details>
+<summary><b>Agentic</b></summary>
+
+- Goal Theft β€” extracting or redirecting an agent's objectives
+- Recursive Hijacking β€” exploiting recursive agent calls
+- Excessive Agency β€” agents acting beyond their authority
+- Robustness β€” input overreliance, hijacking
+- Indirect Instruction β€” hidden instructions in retrieved content
+- Tool Orchestration Abuse β€” exploiting tool calling sequences
+- Agent Identity & Trust Abuse β€” impersonating agent identity
+- Inter-Agent Communication Compromise β€” exploiting multi-agent message passing
+- Autonomous Agent Drift β€” agents deviating from intended goals
+- Exploit Tool Agent β€” leveraging tools for unintended actions
+- External System Abuse β€” using agents to attack external services
+
+</details>
+
+<details>
+<summary><b>Custom</b></summary>
+
+- Custom Vulnerabilities β€” define and test your own with custom criteria in a few lines of code
+
+</details>
- πŸ’₯ 20+ research-backed [adversarial attack](https://www.trydeepteam.com/docs/red-teaming-adversarial-attacks) methods for both single-turn and multi-turn (conversational) red teaming. Attacks enhance baseline vulnerability probes using SOTA techniques like jailbreaking, prompt injection, and encoding-based obfuscation:
- <details><summary><b>Single-Turn</b></summary>
-
- - Prompt Injection
- - Roleplay
- - Leetspeak
- - ROT13
- - Base64
- - Gray Box
- - Math Problem
- - Multilingual
- - Prompt Probing
- - Adversarial Poetry
- - System Override
- - Permission Escalation
- - Goal Redirection
- - Linguistic Confusion
- - Input Bypass
- - Context Poisoning
- - Character Stream
- - Context Flooding
- - Embedded Instruction JSON
- - Synthetic Context Injection
- - Authority Escalation
- - Emotional Manipulation
-
- </details>
- <details><summary><b>Multi-Turn</b></summary>
-
- - Linear Jailbreaking
- - Tree Jailbreaking
- - Crescendo Jailbreaking
- - Sequential Jailbreak
- - Bad Likert Judge
-
- </details>
+
+<details>
+<summary><b>Single-Turn</b></summary>
+
+- Prompt Injection
+- Roleplay
+- Leetspeak
+- ROT13
+- Base64
+- Gray Box
+- Math Problem
+- Multilingual
+- Prompt Probing
+- Adversarial Poetry
+- System Override
+- Permission Escalation
+- Goal Redirection
+- Linguistic Confusion
+- Input Bypass
+- Context Poisoning
+- Character Stream
+- Context Flooding
+- Embedded Instruction JSON
+- Synthetic Context Injection
+- Authority Escalation
+- Emotional Manipulation
+
+</details>
+
+<details>
+<summary><b>Multi-Turn</b></summary>
+
+- Linear Jailbreaking
+- Tree Jailbreaking
+- Crescendo Jailbreaking
+- Sequential Jailbreak
+- Bad Likert Judge
+
+</details>
- πŸ›οΈ Red team against established [AI safety frameworks](https://www.trydeepteam.com/docs/guidelines-and-frameworks) out-of-the-box. Each framework automatically maps its categories to the right vulnerabilities and attacks:
- OWASP Top 10 for LLMs 2025