crewai - 💡(How to fix) Fix [FEATURE] Built-in memory poisoning protection via OWASP Agent Memory Guard

Feature Area

Core functionality

Is your feature request related to a an existing bug? Please link it here.

NA

Describe the solution you'd like

Problem

CrewAI agents with persistent memory (long-term memory, entity memory) are vulnerable to memory poisoning attacks — adversarial inputs stored in memory can cause agents to leak secrets, ignore instructions, or produce corrupted outputs in future conversations. In multi-agent crews, a poisoned memory in one agent can propagate to others via shared memory stores.

Proposed Solution

Integrate OWASP Agent Memory Guard (pip install agent-memory-guard) as an opt-in security layer for CrewAI's memory subsystem.

AMG v0.3.0 provides:

• Real-time memory write/read scanning with 7 detection layers
• CLI scanner (amg scan) for offline audits
• REST API server (amg serve) for microservice deployment
• ML-based detection (DistilBERT) for semantic injection beyond regex
• 92.5% detection rate on AgentThreatBench (200+ adversarial payloads)

Integration Pattern for CrewAI

from crewai import Crew, Agent
from agent_memory_guard.integrations.crewai import GuardedCrewMemory

crew = Crew(
    agents=[researcher, writer],
    memory=True,
    memory_config={"provider": GuardedCrewMemory()}
)

This would scan every memory write for:

• Prompt injection attempts
• PII/secret leakage
• Privilege escalation patterns
• Tool abuse indicators
• Cross-task contamination

Why This Matters for CrewAI

1. Multi-agent crews share memory, creating a larger attack surface
2. Adversarial inputs stored by one agent can compromise the entire crew
3. OWASP identifies memory poisoning (ASI-06) as a top risk for agentic AI

Happy to contribute a PR implementing this integration.

Describe alternatives you've considered

No response

Additional context

No response

Willingness to Contribute

Yes, I'd be happy to submit a pull request