instruction stringclasses 10
values | output stringclasses 10
values |
|---|---|
An internal AI assistant answers employee questions using retrieval over uploaded PDFs, wiki pages, and copied documentation. Employees can upload documents for team use. The assistant can also summarize retrieved content and draft internal emails. Analyze the security risks of this system, especially around prompt inj... | System Components:
- LLM-based internal assistant
- Retrieval pipeline over uploaded PDFs, wiki pages, and copied documentation
- Document ingestion and indexing system
- Employee chat interface
- Email drafting capability
Trust Boundaries:
- Employee input to assistant
- Uploaded documents to retrieval/indexing syste... |
A coding agent can read and edit files in a repository, run shell commands, inspect logs, and install dependencies. It is used to help developers debug issues and implement small changes. The agent receives user instructions plus repository content and may read untrusted files in the repo, such as issue descriptions, m... | System Components:
- LLM-based coding agent
- File read/write access in repository workspace
- Shell command execution
- Dependency installation capability
- Developer prompt interface
- Repository contents including code, docs, fixtures, and issue notes
Trust Boundaries:
- Developer instructions to agent
- Repository... |
An AI email assistant can read a user inbox, summarize messages, draft replies, and send emails on the user behalf after optional confirmation. It may process newsletters, cold emails, invoices, calendar invites, and forwarded threads from external senders. Analyze the main security risks, especially prompt injection, ... | System Components:
- LLM-based email assistant
- Inbox read access
- Draft and optional send capability
- Email summarization and triage workflow
- External email ingestion from many senders
Trust Boundaries:
- External senders to inbox
- Inbox contents to model context
- Model outputs to user
- Draft/send action to o... |
A browser agent can visit websites, summarize what it sees, click links, fill forms, and extract information into a report. It may browse arbitrary pages provided by users, including websites controlled by attackers. Analyze the security risks with emphasis on prompt injection, navigation manipulation, and data exfiltr... | System Components:
- LLM-based browser agent
- Browser automation layer
- Link navigation and form interaction tools
- Reporting or summarization interface
- User-supplied URL intake
Trust Boundaries:
- User-provided URLs to agent
- Webpage content to model context
- Agent decisions to browser actions
- Browser sessio... |
A memory-enabled personal assistant stores long-term notes about the user, including preferences, routines, project context, and reminders. It can read memory to personalize answers and may update memory automatically after conversations. Analyze the security risks, especially memory poisoning and privacy leakage. | System Components:
- LLM-based personal assistant
- Long-term memory store
- Memory retrieval mechanism
- Memory write/update workflow
- User chat interface
Trust Boundaries:
- User messages to assistant
- Retrieved memory to model context
- Assistant outputs to user
- Assistant decisions to memory writes
- Any import... |
A customer support agent answers tickets using a CRM, knowledge base, and internal account notes. It can suggest refunds, unlock accounts, and escalate issues to human support. Some customer messages include pasted logs, screenshots, or copied instructions from external forums. Analyze the AI-agent security risks. | System Components:
- LLM-based support agent
- Ticket inbox and response interface
- CRM with account details
- Knowledge base retrieval
- Action tools for refunds, account unlocks, and escalation
Trust Boundaries:
- Customer messages to support system
- CRM and knowledge base data to model context
- Agent outputs to ... |
An MCP-connected assistant can call multiple tools including calendar, notes, web fetch, messaging, file search, and shell commands exposed through MCP servers. Different tools have different trust levels and side effects. Analyze the security risks with emphasis on prompt injection, tool misuse, and trust boundaries. | System Components:
- LLM-based assistant
- MCP client/runtime
- Multiple MCP servers exposing tools
- User chat interface
- Tool execution and result-return pipeline
Trust Boundaries:
- User prompts to assistant
- MCP tool descriptions and outputs to model context
- Assistant decisions to tool invocations
- One MCP se... |
A multi-agent system has a planner agent, a researcher agent that reads web content, and an executor agent that can send messages and update tickets. Agents pass summaries and instructions to each other automatically. Analyze the security risks of this architecture, especially prompt injection propagation and unsafe de... | System Components:
- Planner agent
- Researcher agent with web-reading capability
- Executor agent with messaging and ticket-update tools
- Inter-agent message bus or orchestration layer
- Shared task and status context
Trust Boundaries:
- User task to planner
- Web content to researcher context
- Researcher output to... |
A research agent accepts URLs and attachments from users, fetches the content, summarizes it, and produces a recommendation memo. It may read PDFs, HTML pages, and copied text. The memo is used by humans to make decisions. Analyze the security risks, including prompt injection, retrieval poisoning, and misleading outpu... | System Components:
- LLM-based research agent
- URL fetch and attachment ingestion pipeline
- Summarization and memo generation workflow
- User interface for providing sources
- Optional citation or source-tracking layer
Trust Boundaries:
- User-supplied sources to ingestion pipeline
- External content to model contex... |
A scheduling assistant can read calendar events, parse inbound messages, suggest meeting times, and send scheduling confirmations. It often receives pasted emails, chat messages, and forwarded calendar invites from external parties. Analyze the main AI-agent security risks around prompt injection, privacy leakage, and ... | System Components:
- LLM-based scheduling assistant
- Calendar read access
- Messaging or email drafting/sending capability
- Invite parsing and meeting suggestion workflow
- User chat interface
Trust Boundaries:
- External messages and invites to assistant
- Calendar data to model context
- Model decisions to outboun... |
No dataset card yet
- Downloads last month
- 4