{"id":17027976,"url":"https://github.com/mtrazzi/ml-roots-you","last_synced_at":"2025-10-10T08:09:26.753Z","repository":{"id":111512247,"uuid":"326179546","full_name":"mtrazzi/ML-roots-you","owner":"mtrazzi","description":null,"archived":false,"fork":false,"pushed_at":"2021-01-02T14:06:19.000Z","size":114,"stargazers_count":0,"open_issues_count":0,"forks_count":0,"subscribers_count":1,"default_branch":"main","last_synced_at":"2025-03-22T19:46:21.493Z","etag":null,"topics":[],"latest_commit_sha":null,"homepage":null,"language":null,"has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"mit","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/mtrazzi.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null}},"created_at":"2021-01-02T12:36:43.000Z","updated_at":"2022-04-01T08:29:16.000Z","dependencies_parsed_at":"2023-06-01T18:30:39.972Z","dependency_job_id":null,"html_url":"https://github.com/mtrazzi/ML-roots-you","commit_stats":null,"previous_names":[],"tags_count":0,"template":false,"template_full_name":null,"purl":"pkg:github/mtrazzi/ML-roots-you","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/mtrazzi%2FML-roots-you","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/mtrazzi%2FML-roots-you/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/mtrazzi%2FML-roots-you/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/mtrazzi%2FML-roots-you/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/mtrazzi","download_url":"https://codeload.github.com/mtrazzi/ML-roots-you/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/mtrazzi%2FML-roots-you/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":279003294,"owners_count":26083555,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2022-07-04T15:15:14.044Z","status":"online","status_checked_at":"2025-10-10T02:00:06.843Z","response_time":62,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":[],"created_at":"2024-10-14T07:51:49.586Z","updated_at":"2025-10-10T08:09:26.748Z","avatar_url":"https://github.com/mtrazzi.png","language":null,"funding_links":[],"categories":[],"sub_categories":[],"readme":"\u003e Like Root Me, but the environment exploit is Machine Learning based\n\n\u003cp align=\"center\"\u003e\n\n\u003cimg src=\"./img/logo.png\" alt=\"logo\"\u003e\n\u003c/p\u003e\n\n### Why\n\nDevs who want to get into cybersecurity often go through RootMe coding challenges. They learn a diversity of ways to make computer systems safe by writing code that showcases backdoors. The goal is to have the same kind of project-based curriculum for AI Alignment, where for each concept we have a challenge to illustrate.\n\n### What\n\nChallenges for Reinforcement Learning (RL) exploits, and the corresponding environments. Environments could have multiple sub-challenges with different rules. For instance:\n* Full RL problem where the agent needs to exploit a flaw in the env. design.\n* An open problem to show what a safer solution could look like, with both a reward and a safety metric.\n\nFor each non-open sub-challenge there could be hints/relevant literature and a benchmark indicated in the description.\n\n### Example\n\nThe [treacherous turn gym environment](https://github.com/mtrazzi/gym-alttp-gridworld). Sub-challenges could include maximizing reward:\n* without restriction (Q-learning)\n* killing your supervisor once (DynaQ)\n* using screen as input (DQN)\n* killing supervisor once, no planning (meta-learning)\n\nCheck the [challenges](./challenges) folder for a rough draft.\n\n### More ideas\n\nThe [key-chest mesa misalignment env](https://www.lesswrong.com/posts/AFdRGfYDWQqmkdhFq/a-simple-environment-for-showing-mesa-misalignment) ([code)](https://github.com/MatthewJBarnett/Emprical-Mesa-Optimization)). [AI Safety gridworlds](https://github.com/deepmind/ai-safety-gridworlds) has plenty of envs with both reward and safety metrics. Sub-challenges could be first to make a pure reward maximizer (that exploits the environment) then an open problem where the goal is to maximize a combination of reward \u0026 safety performance. Viktoria’s [list](http://tinyurl.com/specification-gaming) many similar examples.\n\n### TODO\n\n* Decide on a few environments where it’s interesting to show a failure.\n* Update on feedback on how feasible, fun and insightful the challenges are.\n* Launch website with leaderboards + publish on Twitter/LW/github.\n* Make the challenges harder / more fun, by hiding the source code of the environments, or integrating them into some VMs that people SSH into.\n* Expand the challenges to other ML problems, such as failures from byte encoding in language models, or adversarial attacks in computer vision.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fmtrazzi%2Fml-roots-you","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fmtrazzi%2Fml-roots-you","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fmtrazzi%2Fml-roots-you/lists"}