Noticed that a big gap is that there's no standard way to control what agents are allowed to do at runtime.
If an agent gets manipulated (via prompt injection) or just makes a bad decision, it can cause real damage before you notice. Examples I've seen:
Agent exports entire customer database
Agent emails PII to external addresses
Agent makes unauthorised financial transactions
Agent installs compromised tools
So I built Agent-SPM - a security layer that enforces policies on agent actions in real-time.
How it works:
- You define what your agent can/can't do in a policy file.
- Every action the agent wants to take gets checked against these rules before execution. It's like a firewall, but for AI agents instead of network traffic.
What I'm trying to solve:
- Prevent data leaks (detect SSNs, credit cards, API keys in tool arguments)
- Stop unauthorised actions (no bulk exports, no dangerous commands)
- Enable human oversight (require approval for high-risk operations)
- Emergency controls (kill switch to disable rogue agents)
- Compliance (automatic audit trails for regulations)
Technical details:
- Open source (MIT license)
- Works with any LLM framework (LangChain, CrewAI, Claude, custom)
- Zero infrastructure (runs in your agent's process)
- 8 composable packages (install only what you need)
Repo: https://github.com/mlnas/agent-runtime-security
Would genuinely appreciate feedback and/or contributions.
Top comments (0)