SentinelMind AI

SentinelMind AI: The Agentic Access Control Auditor

💡 Inspiration

Modern cloud applications change rapidly, making business logic and access control boundaries incredibly hard to audit. Security teams struggle with "permission creep" and silent API failures—where a change in code suddenly exposes restricted endpoints (returning an unexpected HTTP 200 instead of a secure HTTP 401).

We wanted to build an autonomous, highly resilient AI security agent that acts as a continuous auditor. To do this, the agent needs a perfect, tamper-proof, and distributed "long-term memory" of the system's security state. That's where we combined the distributed power of CockroachDB with AWS to create SentinelMind AI.

⚙️ What it does

SentinelMind AI is an agentic security companion that maps, monitors, and dynamically audits authorization boundaries.

1. State Profiling: The agent crawls and maps out your API routes, storing the precise response patterns (headers, status codes, and body structures) associated with different privilege levels.
2. Continuous Auditing: It periodically runs non-disruptive, intelligent access control checks.
3. Anomalous Drift Detection: If a configuration change or deployment suddenly exposes sensitive tenant data or permits unauthorized actions, SentinelMind AI detects the deviation by comparing live results against its long-term state memory.
4. Auto-Remediation: It drafts actionable fixes (like AWS IAM policy updates or security group rules) to lock down the leak.

🛠️ How we built it

We engineered a modern, resilient architecture using serverless components:

• The Brain (AI Agent): Built using an agentic framework (such as LangChain/LlamaIndex) to reason through complex access control flows.
• The Agentic Memory (CockroachDB Serverless): We used CockroachDB as our primary distributed state engine. It stores historical access schemas, endpoint metadata, and past vulnerability profiles. Its PostgreSQL compatibility made integration seamless, and its strict serializability ensures the agent's memory of security policies is never corrupted or out of sync globally.
• The Infrastructure (AWS): The entire agent backend is hosted on AWS, using serverless computing to execute audit routines securely and on-demand.

🧠 Challenges we faced

• Handling Distributed State: Traditional vector databases are great for semantic search, but terrible for maintaining strict, transactional security states. We had to learn how to structure relational data in CockroachDB to serve as the structured "long-term memory" for our LLM agents.
• Avoiding False Positives: Distinguishing between an intentional API change and a genuine broken access control vulnerability required fine-tuning the agent's reasoning prompts.

📈 What we learned & What's next

We learned that distributed SQL databases like CockroachDB are uniquely suited for AI agents that require high-reliability memory. For our next steps, we plan to build out-of-the-box integrations for AWS CloudTrail and popular API gateways, allowing SentinelMind AI to ingest real-time traffic logs and block active authorization bypass attempts on the fly.

Built With

• ai-agent
• amazon-web-services
• cockroachdb
• cybersecuirty
• langchain
• postgresql
• python