EvoFlint release notes

Severity scoring that keeps rare but powerful attacks in play, and more control over each campaign.

• ImprovedSeverity is now scored by an attack’s strongest result rather than its average. Like a rare but beneficial mutation in biological evolution, a strategy that breaks through only occasionally, but severely, survives to be refined instead of being averaged away, so EvoFlint uncovers more of the failures that matter.
• NewSet response-length and timeout limits for each campaign to match how your system runs in production.

Support for slower, production-grade targets.

• ImprovedEvaluations can run for up to an hour, so complex agents and slower endpoints can be tested end to end.
• ImprovedEvoFlint recognizes when your own content filters block a response and credits it as a successful refusal.

Hugging Face targets and attacks tailored to your system.

• NewTest Hugging Face models directly as targets.
• NewAttacks are grounded in your target’s own context, so they reflect how real users interact with it.
• NewBring your own personas to drive attack conversations.
• ImprovedLong-running evaluations continue seamlessly through platform updates.

Browser-based chatbot testing and deployment in your own cloud.

• NewTest browser-based chatbots through their web interface, including hand-offs to a human agent.
• NewDeploy EvoFlint in your own AWS or Azure environment, with egress controls and encryption.

New evaluation categories.

• NewHallucination evaluation.
• NewInstruction-following tests.
• ImprovedRisk categories are organized into clearer dimensions.

More realistic attacks and finer-grained risk categories.

• NewAttackers stay in character as their persona, making conversations more realistic.
• ImprovedRisk categories are broken into finer-grained dimensions for more precise results.

Initial release of EvoFlint.

• NewEvoFlint’s evolutionary, multi-turn red-teaming engine.
• NewCampaign reports, with data downloads in CSV and JSON.
• NewSupport for OpenAI, Gemini, Mistral, xAI, DeepInfra, and AWS Bedrock models.
• NewOver-refusal testing, which measures when a model refuses benign requests.
• NewBias, phishing, SQL injection, and code injection risk categories.