--- title: This Open Source Tool Removes AI's Safety Filters description: An open-source tool called Heretic automatically strips safety guardrails from open-weight AI models using a technique called abliteration (directional... canonical: https://daily.dev/posts/this-open-source-tool-removes-ai-s-safety-filters-ghwadlo16 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: This Open Source Tool Removes AI's Safety Filters | daily.dev og:description: An open-source tool called Heretic automatically strips safety guardrails from open-weight AI models using a technique called abliteration (directional... og:url: https://daily.dev/posts/this-open-source-tool-removes-ai-s-safety-filters-ghwadlo16 og:image: https://api.daily.dev/og/posts/GHWaDlO16.png og:image:alt: This Open Source Tool Removes AI's Safety Filters og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # This Open Source Tool Removes AI's Safety Filters **[YouTube](https://daily.dev/sources/youtube)** · 2 min read · 0 upvotes · 0 comments ## Summary An open-source tool called Heretic automatically strips safety guardrails from open-weight AI models using a technique called abliteration (directional ablation), requiring zero fine-tuning or manual tuning. It computes the refusal direction from harmful vs. harmless prompt activations and uses the Optuna library to optimize ablation parameters that minimize both refusal rates and KL divergence from the original model. On a 12-billion parameter Gemma 3 model, it matched manual abliteration's refusal removal rate while achieving a KL divergence of 0.16 versus 1.04, meaning far less damage to the model's reasoning ability. Beyond jailbreaking, the tool is useful for interpretability research, letting researchers visualize how refusal is represented in a transformer's residual stream and test the robustness of alignment training. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is the Heretic tool and how does it remove safety filters from open-weight AI models? Heretic is an open-source tool that automatically strips refusal behavior from open-weight AI models using a technique called abliteration (directional ablation), requiring zero fine-tuning and zero manual effort. It computes the refusal direction as a difference of means between harmful and harmless prompt activations, then uses the Optuna library to search for ablation parameters that minimize both refusals and KL divergence from the original model. _Developers evaluating AI safety tooling can track open-source projects like this one on daily.dev._ ### How much does Heretic's abliteration damage model quality compared to manual abliteration methods? On a 12-billion parameter Gemma 3 model, Heretic achieved the same refusal removal rate as a well-known manual abliteration method but with a KL divergence of only 0.16 versus 1.04, meaning it caused dramatically less damage to the model's reasoning and overall intelligence. _Anyone comparing model-editing techniques can follow benchmark details like these on daily.dev._ ### What research applications does directional ablation (abliteration) have beyond jailbreaking chatbots? Abliteration techniques let researchers study interpretability by mapping how refusal behavior is represented inside a transformer's residual stream, using built-in tooling to plot and analyze those residual vectors. They also allow testing how well alignment training holds up under adversarial pressure, making the technique useful for safety research rather than just circumventing restrictions. _Researchers exploring interpretability techniques can keep up with tools like this via daily.dev._ ## Similar posts on daily.dev - [p-e-w/heretic: Fully automatic censorship removal for language models](https://daily.dev/posts/p-e-w-heretic-fully-automatic-censorship-removal-for-language-models-ycfxbdptu) · Hacker News · 1 upvotes · 0 comments - [Single prompt breaks AI safety in 15 major language models](https://daily.dev/posts/single-prompt-breaks-ai-safety-in-15-major-language-models-ubtbjcjow) · InfoWorld · 0 upvotes · 0 comments --- Tags: [#open-source](https://daily.dev/tags/open-source), [#llm](https://daily.dev/tags/llm), [#ai-safety](https://daily.dev/tags/ai-safety) [View this post on daily.dev](https://daily.dev/posts/this-open-source-tool-removes-ai-s-safety-filters-ghwadlo16) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"This Open Source Tool Removes AI's Safety Filters","url":"https://daily.dev/posts/this-open-source-tool-removes-ai-s-safety-filters-ghwadlo16","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/this-open-source-tool-removes-ai-s-safety-filters-ghwadlo16"},"datePublished":"2026-08-30T19:23:46.692Z","dateModified":"2026-08-30T19:24:08.731Z","description":"An open-source tool called Heretic automatically strips safety guardrails from open-weight AI models using a technique called abliteration (directional...","image":"https://i.ytimg.com/vi/w86c1q59QKU/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/w86c1q59QKU/sddefault.jpg","isAccessibleForFree":true,"articleSection":"YouTube","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"YouTube","logo":"https://media.daily.dev/image/upload/s--W5zMumpP--/f_auto/v1725350181/logos/youtube","url":"https://daily.dev/sources/youtube"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/this-open-source-tool-removes-ai-s-safety-filters-ghwadlo16","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"open-source,llm,ai-safety","timeRequired":"PT2M","video":{"@type":"VideoObject","name":"This Open Source Tool Removes AI's Safety Filters","description":"An open-source tool called Heretic automatically strips safety guardrails from open-weight AI models using a technique called abliteration (directional...","thumbnailUrl":"https://i.ytimg.com/vi/w86c1q59QKU/sddefault.jpg","uploadDate":"2026-08-30T19:23:46.692Z","duration":"PT2M","url":"https://api.daily.dev/r/GHWaDlO16","embedUrl":"https://www.youtube.com/embed/w86c1q59QKU"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"YouTube","item":"https://daily.dev/sources/youtube"},{"@type":"ListItem","position":3,"name":"This Open Source Tool Removes AI's Safety Filters"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/this-open-source-tool-removes-ai-s-safety-filters-ghwadlo16#faq","mainEntity":[{"@type":"Question","name":"What is the Heretic tool and how does it remove safety filters from open-weight AI models?","acceptedAnswer":{"@type":"Answer","text":"Heretic is an open-source tool that automatically strips refusal behavior from open-weight AI models using a technique called abliteration (directional ablation), requiring zero fine-tuning and zero manual effort. It computes the refusal direction as a difference of means between harmful and harmless prompt activations, then uses the Optuna library to search for ablation parameters that minimize both refusals and KL divergence from the original model. Developers evaluating AI safety tooling can track open-source projects like this one on daily.dev."}},{"@type":"Question","name":"How much does Heretic's abliteration damage model quality compared to manual abliteration methods?","acceptedAnswer":{"@type":"Answer","text":"On a 12-billion parameter Gemma 3 model, Heretic achieved the same refusal removal rate as a well-known manual abliteration method but with a KL divergence of only 0.16 versus 1.04, meaning it caused dramatically less damage to the model's reasoning and overall intelligence. Anyone comparing model-editing techniques can follow benchmark details like these on daily.dev."}},{"@type":"Question","name":"What research applications does directional ablation (abliteration) have beyond jailbreaking chatbots?","acceptedAnswer":{"@type":"Answer","text":"Abliteration techniques let researchers study interpretability by mapping how refusal behavior is represented inside a transformer's residual stream, using built-in tooling to plot and analyze those residual vectors. They also allow testing how well alignment training holds up under adversarial pressure, making the technique useful for safety research rather than just circumventing restrictions. Researchers exploring interpretability techniques can keep up with tools like this via daily.dev."}}]} ```