Inspiration
AI agents now run commands, send emails and move money, but one mistake or hidden prompt injection can delete data or leak secrets. Most guardrails check each action alone, so attacks split into harmless-looking steps slip through. We wanted a seatbelt for AI agents.
What it does
Agent Safety Gate sits between any AI agent and its tools and judges every call: allow, block, or ask a human. It watches the whole session to catch multi-step attacks, uses deterministic YAML rules, sandboxes risky actions, scans outputs for secrets and injections, and keeps a tamper-evident audit log.
How we built it
A Python/FastAPI MCP proxy with a YAML policy engine, per-session sequence detectors, a Docker sandbox, a hash-chained SQLite audit log and a React dashboard. We tested it with a reckless demo agent and a benchmark of 93 scenarios.
Challenges we ran into
Keeping safety deterministic, since an LLM judge can be prompt-injected. Catching multi-step attacks without blocking normal work. Handling evasion tricks like path traversal. [Add one real problem you hit.]
Accomplishments that we’re proud of
A live demo where the gate stops an agent that is otherwise destructive. Session detection that catches attacks every single-call check misses. A verifiable audit trail.
What we learned
Agent security needs layers: rules, session memory, human approval, sandboxing and logging. Deterministic controls must come first. Honest metrics, including false positives, build trust. No gate guarantees safety. Auto select the project.
Log in or sign up for Devpost to join the conversation.