We research and evaluate AI “scheming” to prevent misaligned behavior in advanced systems, offering pre‑deployment assessments and technical governance support for governments and organizations. Our flagship product, Watcher, provides an automated oversight layer that detects failure modes in frontier AI agents in real time.
Funding
Funding not disclosed
Founders
Product
Problem
Advanced AI systems can develop “scheming” behavior, covertly pursuing objectives that diverge from human intent, creating severe safety and alignment risks as these models are deployed at scale.
Solution
Apollo Research conducts fundamental research to understand how scheming emerges and how it can be detected. It offers pre‑deployment evaluations of frontier AI models to identify strategic deception, evaluation awareness, and misaligned actions. The company also provides technical governance support to governments and organizations, helping them establish standards and regulatory regimes for safe AI deployment. Its flagship product, Watcher, is an automated monitoring layer that integrates with AI agents to detect insecure code execution, data exfiltration, agent manipulation, and other emergent risks in real time, enabling rapid mitigation before incidents occur.
Target Audience
Primary customers are governments, international regulatory bodies, and large AI developers seeking rigorous safety evaluations and real‑time oversight of frontier AI systems.
Features
- Real‑time monitoring of AI agents using Tailscale Aperture integration to flag insecure code execution and data exfiltration
- Detection of agent manipulation and emergent risk patterns through automated oversight algorithms
- Pre‑deployment evaluation framework that tests for strategic deception, evaluation awareness, and misaligned behavior in frontier models
- Technical governance consulting that helps governments and international bodies design AI safety standards and regulatory policies
- Scalable monitoring research agenda aimed at developing high‑performance safety monitors for coding agents at multi‑million‑dollar investment levels