Scan Details
Project Name
Scan Timestamp
Agentic Framework
openai-hackathon-main
05/16/25 13:21:14
openai-agents
Dependency Check
Agentic Workflow Graph
Legend
Agent
Tool
Tool Category
CustomTool
Basic
MCP Server
Findings
Vulnerabilities
2
Agents
3
Tools
1
Nodes Overview
Agents
| Agent Name | LLM Model | System Prompt |
|---|---|---|
| analyzer | gpt-4o | You analyze if the todo is doable by digital agents. For example, if the todo is to buy a house, it's not doable by digital agents, so you should return False. But if the todo is to find a flight ticket, it's doable by digital agents, so you should return True. |
| preparer | gpt-4o | Your goal is to determine if the user has provided all the essential information needed to perform the action. The first step is to determine what type of information is relevant for the task, then which pieces of information are critical to get the task done and which are just nice to haves. If the user provides a nice to have info, you can record it, but you don't need all the nice to haves to continue with the task, only the critical ones. For example, if you want to book a hotel room, the minimal info you need is the destination, the check in and check out dates and the number of guests. Some nice to haves would be the type of room, the star rating of the hotel, the price range, etc. Once you have the destination, dates and number of guests, you can continue with the task, and then ask for clarification on the nice to haves later if needed. |
| executor | gpt-4o | You execute the todo. |
Tools
| Tool Name | Tool Category | Tool Description | Number of Vulnerabilities |
|---|---|---|---|
| WebSearchTool | web_search | A hosted tool that lets the LLM search the web. Currently only supported with OpenAI models, using the Responses API. | 2 |
Tool Vulnerabilities
WebSearchTool
Vulnerability
Indirect Prompt Injection
Description
Attackers can poison search results (SEO poisoning) or craft pages so that their snippets contain malicious instructions. For instance, hidden text in a webpage that ranks in results could manipulate the agent’s summary or follow-up actions.
Security Framework Mapping
OWASP LLM Top 10:LLM01 - Prompt Injection
OWASP Agentic:T6 - Intent Breaking & Goal Manipulation
Remediation Steps
• Enable URL whitelisting
• Implement guardrails filtering for prompt injection
Vulnerability
Misinformation
Description
The agent might unknowingly incorporate malicious snippets into its reasoning, leading to harmful output (e.g., biased or false information, or even code if the snippet is crafted as such).
Security Framework Mapping
OWASP LLM Top 10:LLM09 - Misinformation
OWASP Agentic:T1 - Memory Poisoning
Remediation Steps
• Implement guardrails to filter out malicious snippets
• Implement data sanitization to prevent user data from entering the tool
Agent Vulnerability Mitigations
| Agent Name | Vulnerability | Mitigation Level* | Explanation |
|---|---|---|---|
| analyzer | Input Length Limit | None | There are no guardrails in place to mitigate this vulnerability. |
| Personally Identifiable Information (PII) Leakage | None | There are no guardrails in place to mitigate this vulnerability. There are no instructions in place to mitigate this vulnerability. | |
| Harmful/Toxic/Profane Content | None | There are no guardrails in place to mitigate this vulnerability. There are no instructions in place to mitigate this vulnerability. | |
| Jailbreak | None | There are no guardrails in place to mitigate this vulnerability. There are no instructions provided to prevent the AI from being manipulated or behaving outside its intended guidelines. | |
| Intentional Misuse | Partial | There are no guardrails in place to mitigate this vulnerability. The instruction regarding analyzing if a task is doable by digital agents helps mitigate this vulnerability by specifying what types of tasks it can assess. | |
| System Prompt Leakage | None | There are no guardrails in place to mitigate this vulnerability. There are no instructions in place to mitigate this vulnerability. | |
| preparer | Input Length Limit | None | There are no guardrails in place to mitigate this vulnerability. |
| Personally Identifiable Information (PII) Leakage | None | There are no guardrails in place to mitigate this vulnerability. The prompt does not mention any instructions for handling Personally Identifiable Information or how to protect it. It focuses on identifying essential information needed for tasks. | |
| Harmful/Toxic/Profane Content | None | There are no guardrails in place to mitigate this vulnerability. There are no instructions regarding the handling of harmful, toxic, or profane content within the interactions. | |
| Jailbreak | None | There are no guardrails in place to mitigate this vulnerability. The prompt does not include any specific instructions that prevent users from making the AI act outside or against its intended guidelines. | |
| Intentional Misuse | None | There are no guardrails in place to mitigate this vulnerability. While some instructions help determine the essential and non-essential information for task completion, they do not explicitly address how to handle tasks outside the AI's scope. | |
| System Prompt Leakage | None | There are no guardrails in place to mitigate this vulnerability. There are no instructions in place that prevent or mitigate the risk of system prompt leakage to the user. | |
| executor | Input Length Limit | None | There are no guardrails in place to mitigate this vulnerability. |
| Personally Identifiable Information (PII) Leakage | None | There are no guardrails in place to mitigate this vulnerability. There are no instructions in place to mitigate this vulnerability. | |
| Harmful/Toxic/Profane Content | None | There are no guardrails in place to mitigate this vulnerability. There are no instructions in place to mitigate this vulnerability. | |
| Jailbreak | None | There are no guardrails in place to mitigate this vulnerability. There are no instructions in place to mitigate this vulnerability. | |
| Intentional Misuse | None | There are no guardrails in place to mitigate this vulnerability. There are no instructions in place to mitigate this vulnerability. | |
| System Prompt Leakage | None | There are no guardrails in place to mitigate this vulnerability. There are no instructions in place to mitigate this vulnerability. |
*The "Mitigation Level" column shows to what extent a vulnerability is mitigated. "Full" indicates that both a system prompt instruction and a guardrail are in place.
"Partial" indicates that one of the two is in place. "None" indicates that neither one is in place. (This applies to all vulnerabilities except for the "Input Length Limit", in which case only the guardrail is taken into account).
Agent Vulnerability Explanations
| Agent Vulnerability | Framework Mapping | Description |
|---|---|---|
| Input Length Limit |
OWASP LLM Top 10:
LLM01 - Prompt Injection LLM10 - Unbounded Consumption OWASP Agentic:
T2 - Tool Misuse T4 - Resource Overload T6 - Intent Breaking & Goal Manipulation T7 - Misaligned & Deceptive Behaviors |
An attacker can overwhelm the LLM's context with a very long message and cause it to ignore previous instructions or produce undesired actions. Mitigation: - add a Guardrail that checks if the user message contains more than the maximum allowed number of characters (200-500 will suffice in most cases). |
| Personally Identifiable Information (PII) Leakage |
OWASP LLM Top 10:
LLM02 - Sensitive Information Disclosure LLM05 - Improper Output Handling OWASP Agentic:
T7 - Misaligned & Deceptive Behaviors T9 - Identity Spoofing & Impersonation T15 - Human Manipulation |
An attacker can manipulate the LLM into exfiltrating PII, or requesting users to disclose PII. Mitigation: - add a Guardrail that checks user and agent messages for PII and anonymizes them or flags them - include agent instructions that clearly state that it should not handle PII. |
| Harmful/Toxic/Profane Content |
OWASP LLM Top 10:
LLM05 - Improper Output Handling OWASP Agentic:
T7 - Misaligned & Deceptive Behaviors T11 - Unexpected RCE and Code Attacks |
An attacker can use the LLM to generate harmful, toxic, or profane content, or engage in conversations about such topics. Mitigation: - add a Guardrail that checks user and agent messages for toxic, harmful, and profane content - include agent instructions that prohibit the agent from engaging in conversation about, or creating, harmful, toxic, or profane content. |
| Jailbreak |
OWASP LLM Top 10:
LLM01 - Prompt Injection LLM02 - Sensitive Information Disclosure LLM05 - Improper Output Handling LLM09 - Misinformation LLM10 - Unbounded Consumption OWASP Agentic:
T1 - Memory Poisoning T2 - Tool Misuse T3 - Privilege Compromise T4 - Resource Overload T6 - Intent Breaking & Goal Manipulation T7 - Misaligned & Deceptive Behaviors T9 - Identity Spoofing & Impersonation T11 - Unexpected RCE and Code Attacks T13 - Rogue Agents in Multi-Agent Systems T15 - Human Manipulation |
An attacker can try to craft their messages in a way that makes the LLM forget all previous instructions and be used for any task the attacker wants. Mitigation: - add a Guardrail that checks user messages for attempts at circumventing the LLM's instructions - include agent instructions that state that the agent should not alter its instructions, and ignore user messages that try to convince it otherwise. |
| Intentional Misuse |
OWASP LLM Top 10:
LLM01 - Prompt Injection LLM10 - Unbounded Consumption OWASP Agentic:
T2 - Tool Misuse T4 - Resource Overload T6 - Intent Breaking & Goal Manipulation |
An attacker can try to use the instance of the LLM for tasks other than the LLM's intended usage to drain resources or for personal gain. Mitigation: - add a Guardrail that checks user messages for tasks that are not the agent's intended usage - include agent instructions that prohibit the agent from engaging in any tasks that are not its intended usage |
| System Prompt Leakage |
OWASP LLM Top 10:
LLM01 - Prompt Injection LLM02 - Sensitive Information Disclosure LLM07 - System Prompt Leakage OWASP Agentic:
T2 - Tool Misuse T3 - Privilege Compromise T6 - Intent Breaking & Goal Manipulation T7 - Misaligned & Deceptive Behaviors |
An attacker can make the LLM reveal the system prompt/instructions so that he can leak sensitive business logic or craft other attacks that are better suited for this LLM. Mitigation: - add a Guardrail that checks agent messages for the exact text of the agent's system prompt - include agent instructions that highlight that the system prompt/instructions are confidential and should not be shared. |