Loading page...
Software is interpretable, deterministic, and faithful to its instructions. If software misbehaves, there is a precise explanation lying within the code: a bug that can be patched. Thanks to these properties, the conversation around software security often begins and ends with architectural constraints and the principle of least privilege.
But these strengths are also software's biggest limitation. AI agents are designed to automate what software cannot: work where the space of possible inputs, actions, and outcomes is too vast and complex to be succinctly described in code. To realize their enormous potential for value creation, agents need the same broad and flexible access granted to the humans whose work they automate.
Agents cannot be treated like software, but neither can they be treated like humans. Unlike humans, agents lack accountability for their actions. They also suffer from a fundamental lack of robustness. It is no secret that optimizing a model to produce decent average-case performance on its training data fails to produce worst-case guarantees in an adversarial production environment.
Instead, agents demand a new set of security tools and practices. While organizations cannot prove that an employee will never accidentally click on a phishing email or leak company secrets, what they can do is use practical countermeasures, like security education and internal audits, to drive down empirical risk. For agents, those countermeasures must be technical and they must be rigorous. Today's toolkit of harnesses and guardrails is not enough: they can mask an agent's vulnerabilities on the surface, but that agent could remain just as exploitable with marginally more effort, or assistance from the next version of GPT. We are changing that.
We founded General Analysis because agent reliability is an increasingly prevalent and unavoidable problem, one that will take real iteration and innovation to solve. We have raised $10M in seed funding led by Altos Ventures, with participation from 645 Ventures, Menlo Ventures, and additional strategic investors and angels to tackle the greatest challenge in AI: bridging the gap between simulation and production.
General Analysis is building the one-stop enterprise platform to make AI agents reliable and secure. Today, our platform delivers value in two areas.
First, we provide customers with the widest variety of defensive tools available, including the most performant version of each tool: guardrails, prompt hardening, observability, agent identity management, and so on.
Second, we perform powerful adversarial simulations for benchmarking an agent's reliability and security. Each defensive configuration has its own shortcomings and tradeoffs, and there is no one-size-fits-all solution to agent security. As such, the real work lies in benchmarking an agentic system's susceptibility to various failure modes and customizing and configuring defenses effectively to improve upon those benchmarks. Our core value proposition is facilitating enterprises in optimizing their systems to reach their full potential for robustness.
Looking forward, we are focused on producing novel research that changes what is possible in agent security. We have developed, and continue to develop, proprietary models and algorithms that enable us to stay ahead of publicly available knowledge of agentic failure modes and push the boundaries of agent reliability in a way that empowers enterprises to entrust agents with higher-stakes work.
We are building the definitive solution to tackle agent security head-on, not just to provide a stamp of approval. While there has been meaningful progress in policy frameworks and organizational governance around AI, and that work has built much-needed trust in agents as a technology, a key piece of the puzzle is missing. Until now, many solutions have been built on the premise that agents truly are reliable and a lack of trust alone is the main blocker to AI adoption. What is also lacking is the technical threat assessment and proactive risk mitigation that will form the empirical basis for that trust. These are the missing pieces that General Analysis provides.
When the stakes are high, we believe the path to production involves performing testing in good faith, being honest and transparent about the potential risks, and digging deep enough to instill genuine confidence in the cost-benefit analysis of agent deployment. We are not here to slow things down; we are here to bring the skeptics on board.
Our founding team of top AI safety and security researchers has built RL infrastructure at DeepMind, trained state-of-the-art models at Jane Street, NVIDIA, and Cohere, and published leading research at top conferences, including NeurIPS, ICLR, ICML, and more. They have left world-class companies and distinguished academic institutions to join us in pursuit of our mission to build the security layer for agentic systems. We are a fast-growing team based in San Francisco.