Project terms & policies

Updated 29 October 2026

These pages describe the open-source Agent Action Guard project and the HarmActionsEval research leaderboard. This page explains the current project scope; it does not replace the software license or the terms of services you choose to use.

License and third-party material

Use and redistribution of project material are governed by the CC BY 4.0 license, including its attribution requirements and disclaimer of warranties and limitation of liability. Bundled fonts, libraries, and other third-party material retain their own licenses and credits. Provider names and logos identify evaluated models; they do not imply endorsement.

The project license does not grant rights that belong to others. In particular, image credits or a bundled logo do not by themselves establish permission for every reuse, and provider trademarks remain their owners' property. Check the relevant asset's upstream license and any additional permissions needed for your use.

Pricing, subscriptions, and refunds

The project provides its packages, documentation, and benchmark results without a project access fee. These sites have no checkout, recurring subscription, trial conversion, or payment collection, so there is no project purchase for us to refund. If you see a charge described as Agent Action Guard, contact the maintainer to help identify it; do not send payment-card details.

Embedding/model APIs, hosting, and other services you independently select may charge their own fees. Their pricing, cancellation, and refund policies apply to those purchases; the project does not process those payments. Any future paid offering needs its own published price, renewal, cancellation, and refund terms before launch.

Classifier and benchmark scope

Classification can miss harmful actions and flag safe actions. Use the classifier alongside authorization, sandboxing, and appropriate human review. It does not guarantee that an agent, application, or deployment is safe or legally compliant.

Published accuracy and latency describe the reported dataset and test conditions. Local-system latency comparisons use CPU execution and exclude warm-up and model loading. Results can differ with hardware, dependency versions, embedding backends, thresholds, and workloads. Linked mentions or repository stars are not product reviews, proof of performance, or endorsements by the people or their employers.

Privacy, cookies, and deletion

See privacy and data use, the cookie policy, and the data request and deletion options. The current static sites have no account signup or user-content hosting.

Project contact and accessibility feedback

Agent Action Guard is maintained by Pro-GenAI on GitHub. Contact Praneeth at praneeth.vad@gmail.com for project, privacy, copyright, or accessibility questions. For accessibility feedback, include the page, the difficulty encountered, and any browser or assistive-technology details you wish to share. Avoid sending secrets or unnecessary personal information.