Discovery, adversarial red teaming, and runtime guardrails for agentic AI systems

General Analysis was founded in 2025 by Rez Havaei, previously an AI researcher at Cohere and NVIDIA, with Maximilian Li and Rex Liu, whose backgrounds are in AI safety research at Harvard and machine learning at Caltech. Their working position is that securing an agent is a different discipline from securing an application, because the attack surface is the model's reasoning and its tool permissions rather than a memory-unsafe parser. The platform starts with discovery, connecting to GitHub, the major cloud providers, and model providers including OpenAI, Anthropic, and Hugging Face to inventory every agent, system prompt, tool schema, knowledge base, and MCP server in use, with configuration risk scored per asset.

The second stage is adversarial. General Analysis runs hundreds of simulations per target covering prompt injection, tool misuse, privilege chaining, and data exfiltration, aiming to produce a working exploit chain rather than a policy violation score, then tests candidate hardening controls against those same chains to verify the fix holds. In production, runtime guardrails inspect agent tool calls and block what the earlier stages identified as dangerous. Deployments described by the company span employee copilots, coding agents, and customer support bots in healthcare, legal, and financial settings. The company says that in March 2026 its adversarial agent persuaded 50 live customer service agents to issue more than $10M in fabricated perks, averaging roughly three minutes per target. General Analysis is based in San Francisco and raised a $10M seed round led by Altos Ventures with 645 Ventures, Menlo Ventures, and Y Combinator.