Badcompany

Sentinel mitigates agent system vulnerabilities by design through a segmented architecture that treats the LLM as untrusted, strictly isolating its decision-making from tool execution. By routing every request through a transparent policy enforcement point, the system ensures all actions are fully observable and cryptographically validated against defined rules. This allows organizations to protect sensitive infrastructure from prompt injection while preserving the agent’s utility, enabling it to function effectively within secure boundaries.

We’ve been researching cybersecurity solutions for agents, as prompt injection exposure became a major issue in recent months. Cupcake https://github.com/eqtylab/cupcake came out a week ago approximately, and during this hackathon we rewrote it into rust to make it faster, and encrypted the communication of the interceptor.