Products/AI Security / Agent Security/Benchspan

Benchspan

Real-Time Security for AI Agents in Production

AI Security / Agent SecurityBacked byFounded 2026Real-time inline threat blocking on request pathIndirect prompt injection detectionData exfiltration preventionUnauthorized tool access blockingCustom model training from your trafficThreat review queue with confirmed alertsCompliance reporting (EU AI Act, NIST AI RMF, OWASP)Behavioral drift detectionAnomalous tool call sequence detection
Visit Benchspan →
Benchspan

Our Take

Benchspan is low-key solving the AI security problem everyone else is ignoring — indirect prompt injection, where attacks hide inside the documents, emails, and tool outputs that agents actually read while working. Their detector specifically trained for this (not generic prompt injection) nailed a 99.9% catch rate on AgentDojo versus 71% for the next best commercial option, and they've already flagged 23 threats with 7 confirmed across 34 monitored agents, which is the kind of traction that makes platform companies pay attention. The 14ms average latency means it actually works inline without turning your agent pipeline into a mess, and the compliance reporting covers EU AI Act and NIST AI RMF for the enterprise crowd, which is the move if you're selling to agent platform teams in 2026.

A custom security model built for AI agents that stops attacks your guardrails miss. It sits on the request path and evaluates every LLM call, tool invocation, and RAG retrieval in real-time.

Problem It Solves
Indirect prompt injection attacks - attacks hidden inside emails, documents, tool responses, and web pages that agents read while doing their job. Every major agent platform (Microsoft, OpenAI, Google, Salesforce, GitLab, Perplexity) has had a zero or one-click indirect prompt injection breach publicly disclosed in the last twelve months.
Target Customer
AI agent platform companies running agents in production
Use Cases
Protect AI agents from prompt injection attacks, Prevent data exfiltration from agents, Block unauthorized tool access, Generate compliance evidence for auditors, Monitor agent behavior and threats
Differentiator
The only detector specifically built for indirect prompt injection attacks that target agents. Trained on attacks hidden in tool outputs, APIs, and web pages - not direct user-to-chatbot attacks. 99.9% catch rate on AgentDojo vs 71% for next best commercial detector.
Why Now
Every major agent platform has had a zero or one-click indirect prompt injection breach publicly disclosed in the last twelve months. Microsoft, OpenAI, Google, Salesforce, GitLab, Perplexity - all in 2025.
Traction
Notable Metrics: 34 agents monitored, 5 environments, 84,291 traces in 24h (+12% vs yesterday), 23 threats flagged (7 confirmed), Avg latency 14ms, P99 latency 42ms, 847 allowed requests, 23 blocked requests · Press Mentions: The Hacker News - Zero-Click AI Vulnerability Exposes Microsoft 365 Copilot Data (June 2025)

Key Facts

Category
AI Security / Agent Security
Location
, United States
Founded
2026
Stage
Backed by
Pricing
Not explicitly mentioned on page
Discovered via
yc

The people behind Benchspan

A

Adam

profile
A

Anant

profile
C

Chase R. W. Johnson

profile
R

Rahul Raghavan

profile

Links

Website

Want products like this in your inbox every morning?

Five products. Every morning. Written by someone who actually cares whether they're good or not. Free forever, unsubscribe whenever.