Ian Webster

Ian Webster co-founded Promptfoo with Michael D’Angelo to help developers test and secure AI applications. He designed its first version while leading AI engineering at Discord; the project developed from prompt evaluation into adversarial testing of application behavior. In March 2026, the two founders announced an agreement for Promptfoo to join OpenAI.[3][4][6]

In his July 2024 account, Webster described starting Discord’s Platform Ecosystem organization and spending years building developer APIs before leading a team developing LLM products for millions of users. He recalled that ensuring safe, secure, reliable behavior was the hardest part of shipping those products, and that conventional testing and security methods were ineffective against the breadth of possible LLM behavior. He designed Promptfoo for application developers to test, discover, and fix those failures.[3]

D’Angelo, a longtime friend with machine-learning leadership experience at Smile Identity, joined shortly afterward. Together they developed a simple, declarative open-source toolkit for evaluating prompts and models. Promptfoo’s first open-source release was in 2023; Webster and D’Angelo founded the company in 2024.[3][4][6]

Webster’s central argument was that AI security should focus on the application layer, “where the model meets the real world.” In his July 2024 announcement, he distinguished model-level exploits from failures specific to an application: customer-service agents revealing database information, assistants making unauthorized promises, or tutors departing from their intended role. He emphasized context poisoning, tool misuse, and use-case hijacking as risks developers needed to test for.[3]

The move toward red teaming was a collective decision. In his June 2025 retrospective, D’Angelo explained that testing expected successful behavior was insufficient: teams also needed to discover vulnerabilities proactively. Promptfoo therefore invested in adversarial testing that simulated malicious inputs, including prompt injections and toxic content, so developers could find weaknesses before deployment.[4]

In July 2025, Webster and D’Angelo extended that argument to tool-using agents. They highlighted agents’ access to production systems, retrieval manipulation that could expose RAG knowledge bases, and injection vectors introduced by Model Context Protocol (MCP) servers. Their proposed roadmap included tracing behavior across multiple steps, capturing and replaying tool calls, and scoring those calls for correctness and security. Their longer-term vision was to evaluate and red-team every model, prompt, or tool change automatically, then block or promote it with a signed safety report.[5]

Webster also argued that security tools should follow open-source principles of “transparency and interoperability” rather than proprietary lock-in. When he and D’Angelo announced the OpenAI acquisition agreement on March 9, 2026, they said joining would let them integrate their security and evaluation technology into model and infrastructure layers. They committed to maintaining the open-source suite and supporting diverse providers and models. The announcement made closing subject to customary conditions; integration was described as future work.[3][6]

Before his AI work, Webster created Asterank, an asteroid database and visualization project acquired by Planetary Resources in May 2013. It provides earlier context for his work building software around scientific data.[2][7]

Asterank estimates potential asteroid value using mass, composition inferred from spectral type, and market prices. Its accessibility estimates emphasize delta-v—the required change in velocity—and orbital characteristics; mining costs are modeled as a flat percentage of potential value. These are estimates with substantial limitations: most asteroids lack spectral classifications and other important data, preventing complete estimates of value or mining cost. Webster’s Asterank code was also generalized into SpaceKit, a library for interactive 3D space visualizations.[7]