promptfoo 0.125.0 adds eval safe mode, new red-team encoding strategy, and rouge-l/s assertions

promptfoo 0.125.0 (Oct 23) is a breaking release that makes jks-js an optional peer dependency and Claude Agent/Codex SDKs opt-in. Notable security content: an 'eval safe mode' to disable inline JavaScript execution, a new 'arabic-presentation-forms' static encoding red-team strategy, rouge-l/rouge-s assertion types, and a Daybreak access-program provider config.

A security-relevant hardening inside a widely used LLM-eval/red-teaming OSS tool: disabling inline JS execution in eval reduces a prompt-injection-style code-execution surface, while the new encoding strategy extends red-team coverage to bypass-and-bypass detection (encoding-based attacks).

Teams running promptfoo for LLM evaluation and red-teaming should pin/upgrade and review the breaking change and enable the new safe-mode and encoding strategy for prompt-injection testing.