Security: OWASP Agent Memory Guard integration for memory poisoning defense

Feature Request: Memory Security for OpenHands Agents

Hi OpenHands team,

OpenHands (formerly OpenDevin) is one of the most capable open-source software engineering agents. As agents operate with increasing autonomy — reading files, executing code, browsing the web — the risk of memory poisoning attacks grows significantly.

The threat: A malicious file, webpage, or tool output can inject content into the agent's memory/context that biases all subsequent decisions — a form of persistent prompt injection.

OWASP Agent Memory Guard is the reference implementation for OWASP ASI06 (Memory Poisoning):

• Runtime scanning of memory reads/writes for injection payloads and structural tampering
• SHA-256 integrity baselines to detect post-write tampering
• Policy-driven enforcement (block / warn / strip)
• Sub-100μs overhead — zero impact on agent performance

pip install agent-memory-guard

GitHub: https://github.com/OWASP/www-project-agent-memory-guard
PyPI: https://pypi.org/project/agent-memory-guard/

Would love to explore adding this as a security layer in OpenHands' memory/context management. Happy to contribute a PR.

No repository files, tests, or entry points are named. Start by locating OpenHands' memory/context management and reviewing the linked OWASP Agent Memory Guard package; done would require an agreed integration scope and validation of its runtime scanning, integrity, and policy behavior.