{"id":115,"date":"2026-10-29T07:02:57","date_gmt":"2026-10-29T07:02:57","guid":{"rendered":"https:\/\/bugskan.com\/?p=115"},"modified":"2026-10-29T07:02:57","modified_gmt":"2026-10-29T07:02:57","slug":"poisoning-study-found-250-documents-could-trigger-a-narrow-model-backdoor","status":"publish","type":"post","link":"https:\/\/bugskan.com\/index.php\/2026\/10\/29\/poisoning-study-found-250-documents-could-trigger-a-narrow-model-backdoor\/","title":{"rendered":"Poisoning study found 250 documents could trigger a narrow model backdoor"},"content":{"rendered":"<p>Anthropic, the UK AI Security Institute and the Alan Turing Institute studied training-data poisoning in models ranging from 600 million to 13 billion parameters. In their setup, 250 malicious documents could teach a trigger that caused gibberish output.<\/p>\n<p>The experiment challenges the assumption that a poisoning attack must control a fixed percentage of an expanding training corpus. Its demonstrated behavior was deliberately narrow, rather than theft of sensitive information or arbitrary code execution.<\/p>\n<p>The researchers said it remained unclear whether the pattern extends to larger models or more harmful behaviors. Treat the result as evidence for stronger data provenance and further evaluation, not proof that every production model can be compromised by the same recipe.<\/p>\n<p><a href=\"https:\/\/www.anthropic.com\/research\/small-samples-poison\">Read the original source<\/a>.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>The result covered gibberish-output triggers in experimental models, with important limits on what it proves about frontier systems.<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"bugskan_incident_id":"inc_4f27b3a14168b3","bugskan_event_type":"research","bugskan_response":"unknown","bugskan_technical_severity":"","bugskan_technical_severity_source":"","bugskan_evidence_quality":"high","bugskan_affected_scope":"Experimental training-data poisoning in 600M\u201313B parameter models.","bugskan_recommended_next_step":"Review training-data provenance and evaluate poisoning risks within your actual model pipeline.","bugskan_who_should_check":"","bugskan_material_updated_at":"2025-10-09T00:00:00Z","bugskan_reviewer_name":"","bugskan_reviewed_at":"","bugskan_legacy_state":"mapped","bugskan_featured":false,"bugskan_sample":false,"footnotes":""},"categories":[1],"tags":[],"bugskan_format":[3],"bugskan_technology":[],"class_list":["post-115","post","type-post","status-publish","format-standard","hentry","category-uncategorized","bugskan_format-research-explained"],"_links":{"self":[{"href":"https:\/\/bugskan.com\/index.php\/wp-json\/wp\/v2\/posts\/115","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/bugskan.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/bugskan.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/bugskan.com\/index.php\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/bugskan.com\/index.php\/wp-json\/wp\/v2\/comments?post=115"}],"version-history":[{"count":1,"href":"https:\/\/bugskan.com\/index.php\/wp-json\/wp\/v2\/posts\/115\/revisions"}],"predecessor-version":[{"id":340,"href":"https:\/\/bugskan.com\/index.php\/wp-json\/wp\/v2\/posts\/115\/revisions\/340"}],"wp:attachment":[{"href":"https:\/\/bugskan.com\/index.php\/wp-json\/wp\/v2\/media?parent=115"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/bugskan.com\/index.php\/wp-json\/wp\/v2\/categories?post=115"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/bugskan.com\/index.php\/wp-json\/wp\/v2\/tags?post=115"},{"taxonomy":"bugskan_format","embeddable":true,"href":"https:\/\/bugskan.com\/index.php\/wp-json\/wp\/v2\/bugskan_format?post=115"},{"taxonomy":"bugskan_technology","embeddable":true,"href":"https:\/\/bugskan.com\/index.php\/wp-json\/wp\/v2\/bugskan_technology?post=115"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}