Manufacturing Intelligence: AI Poisoning Demands Pharma Data Stewards

In the latest episode of Manufacturing Intelligence, PharmTech Associate Editorial Director Chris Cole and cohost Richard Jaenisch, senior director of Education, Outreach, and Digital Experience at Open Biopharma, move from diagnosing AI poisoning to defending against it. The answer is less about new software and more about a new role, the data steward.

Jaenisch describes the threat as contamination of the information supply chain. Every organization relies on a stream of trusted inputs, such as specific publications, institutions, and individuals. A manufacturer already protects its clean rooms from contamination, and it now has to protect the instructions and memory inside its AI tools in the same way. The data steward doesn't need to carry that title, he explains. The function could sit in several places, though it most likely belongs in an existing data governance structure. That person watches "the memory, the instructions, all of these weird little pieces" in externally exposed AI implementations and machine learning models. They monitor for drift that signals outside influence.

Off-the-shelf tools may not close the gap. Jaenisch notes that a product like Copilot probably doesn't check for this kind of manipulation. A recent OpenAI security offering seems to catch some data poisoning but has issues with certain kinds of memory poisoning, because a malicious instruction can read like an ordinary sentence. Reviewing instructions may also require permission trees that current software doesn't support.

His remedy is an immutable ledger that records what each instruction set should say, so stewards can compare it against what exists. At scale, that takes automation. Manual review is impractical in a 20,000-person company where half the staff use AI assistants with memory functions.

The role doesn't require a coding background. "This isn't about the code," Jaenisch says. "This is about the verbiage. It's about the linguistics." Hidden terms buried in long instruction sets are easy to miss, which is why, he adds, an additional AI assistant could help.