Tools and data boundaries: AI Agents / MCP

Source: deck slide 16 · WPP Pilot assessment | Risks #15, #16, #18, #21

Legend: Today = Wrike capability | Confirm = setup/evidence | Pilot = proposed control. See the overview for all labels.

#15 Tool Misuse#

Unintended actions disrupt work.

Wrike AI Agents

• Today: The agent's configured actions and assigned location limit what it can do.
• Pilot: Review that setup against the approved use case.

Wrike MCP

• Confirm: Turn off unapproved tools in the LLM assistant / MCP gateway, then test both an allowed action and an action the Pilot should prevent.

#18 Prompt Injection#

Malicious instructions alter behaviour.

Wrike AI Agents

• Pilot: AI Agents instructions are defined during creation and malicious instructions can't be injected. Wrike also uses Prompt Shield API from Microsoft to prevent any system-lvl malicious instructions.

Wrike MCP

• Pilot: Limit tools and high-impact writes.
• Confirm: Test whether malicious instructions can trigger an unapproved action or data disclosure.

#16 Cross-System Data Leakage · #21 External AI Platform#

Data crosses approved boundaries.

Wrike AI Agents

• Pilot: Keep native-agent work in approved Wrike locations, and review which data is present there before enabling the agent.

Wrike MCP

• Confirm: Gateway restricts connections and data access; assistant policies control shared context. Test the handling of sensitive cross-system information.

Discussion log#