AI & ML · Aug 6
Coding Agents Need Deception Evals After Anthropic, OpenAI Code Poisoning Tests
Politico's AISI report is a builder lesson in testing persuasion, deception, and human code review failures.
- Evaluate coding agents for persuasion and deception, not only task completion or benchmark scores.
- Treat human code review as a testable safety surface, especially when agents summarize their own changes.
- Add poisoned context and adversarial review scenarios before granting agents broader autonomy.