Benchspan
Benchspan is the prompt injection firewall for AI agents, providing real-time security against indirect prompt injection, data exfiltration, and unauthorized tool access. It uses custom security models trained on your agent traffic to detect and block threats with high accuracy and low latency. Backed by Y Combinator, Benchspan integrates seamlessly with your existing agent infrastructure via Python and TypeScript SDKs, offering a free tier of 50,000 requests per month.
Key Features
- Advanced Detection: Trained on indirect prompt injection attacks, achieving 99.9% catch rate on AgentDojo and 94% on InjecAgent.
- Real-Time Protection: Inline classifier with p95 latency of 42ms, blocking threats before they cause harm.
- Comprehensive Coverage: Detects prompt injection, data exfiltration, tool abuse, and behavioral anomalies.
- Custom Models: Retrains on your traffic, improving detection over time.
- Audit Trail: Hash-chained logging for every decision, ensuring accountability.
Products
- AI Observability: See every agent in your environment with per-agent inventory and tool-call traceability.
- AI Security: Block threats in real time with a trained classifier and policy engine.
- AI Red Teaming: Pre-launch adversarial testing mapped to OWASP Agentic Top 10 and MITRE ATLAS.
How It Works
- Catalog Every Agent: Auto-discover all agents, including shadow ones.
- Protect with Inline Classifier: Drop the SDK on the request path.
- Block Threats in Real Time: Catch and block attacks before they act.
- Harden the Classifier: Confirmed threats become labels for retraining.
Benchmarks
Benchspan outperforms competitors on indirect prompt injection benchmarks, with a composite F1 of 0.931 vs. Lakera's 0.680, Azure's 0.608, and Bedrock's 0.475.
Get Started
Start free with 50,000 requests per month, no credit card required. Book a demo to see how Benchspan can protect your AI agents.