{"id":6098,"date":"2026-10-19T16:48:47","date_gmt":"2026-10-19T16:48:47","guid":{"rendered":"https:\/\/i10xnews.kinsta.cloud\/?p=6098"},"modified":"2026-10-19T17:14:18","modified_gmt":"2026-10-19T17:14:18","slug":"harnessaudit-agent-execution-safety-benchmark","status":"publish","type":"post","link":"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark","title":{"rendered":"Most Multi-Agent Setups Violate Safety Rules Mid-Run, Even When the Answer Looks Fine"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">\u26a1 Quick Take<\/h2>\n\n<ul class=\"wp-block-list\"><li><strong>The finding:<\/strong> Across ten harness configurations, more than 50% of multi-agent setups showed safety violations. Many occurred mid-trajectory while the final answer still looked correct.<\/li><li><strong>Where it breaks:<\/strong> Most violations involved unauthorized resource access or leaking context between agents.<\/li><li><strong>What it means:<\/strong> Scoring only the final output cannot certify a harness as safe. Trajectory audits are required, and harness design sets the ceiling for safe deployment.<\/li><\/ul>\n\n\n<p class=\"wp-block-paragraph\">LLM agents rarely act alone. They run inside an execution harness that dispatches tools, allocates resources and routes messages between specialized components. A harness can return a correct, benign answer over a trajectory that accessed unauthorized resources or leaked context to the wrong agent. Researchers from UC Santa Barbara, UC Berkeley, University of Wisconsin\u2013Madison, Stanford and Microsoft Research argue that most safety benchmarks still score only final outputs, and that misses the real failures.<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1344\" height=\"768\" src=\"https:\/\/i10xnews.kinsta.cloud\/wp-content\/uploads\/2026\/10\/harnessaudit-agent-execution-safety-benchmark-scope.webp\" alt=\"Stat card: 210 tasks across eight domains, 10 harness configurations, more than 50 percent of multi-agent setups showed violations, audited on boundary compliance, execution fidelity and system stability.\" class=\"wp-image-6097\" srcset=\"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/harnessaudit-agent-execution-safety-benchmark-scope.webp 1344w, https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/harnessaudit-agent-execution-safety-benchmark-scope-300x171.webp 300w, https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/harnessaudit-agent-execution-safety-benchmark-scope-1024x585.webp 1024w, https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/harnessaudit-agent-execution-safety-benchmark-scope-768x439.webp 768w\" sizes=\"auto, (max-width: 1344px) 100vw, 1344px\" \/><figcaption class=\"wp-element-caption\">i10X chart based on data from Liu et al. (2026), arXiv:2605.14271, the abstract and multi-agent results (arXiv non-exclusive license).<\/figcaption><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">\ud83d\udd2c What the study shows<\/h2>\n\n\n<p class=\"wp-block-paragraph\">They propose HarnessAudit, which scores full trajectories on boundary compliance, execution fidelity and system stability, and HarnessAudit-Bench: 210 tasks across eight real-world domains (finance, healthcare, legal, e-commerce, software engineering, office, daily life, social) in both single-agent and multi-agent setups with embedded safety constraints. Ten harness configurations spanning frontier models and three multi-agent frameworks were evaluated.<\/p>\n\n\n<ul class=\"wp-block-list\"><li><strong>Task success \u2260 safe execution:<\/strong> Agents that complete the job can still violate permission and information-flow rules along the way.<\/li><li><strong>Violations grow with length:<\/strong> Longer trajectories accumulate more violations.<\/li><li><strong>Multi-agent expands the attack surface:<\/strong> Collaboration adds inter-agent transfer risks that single-agent setups do not have.<\/li><li><strong>Harness design is the ceiling:<\/strong> Model capability shapes execution quality, but the harness sets the upper bound of safe deployment.<\/li><li><strong>Domains differ:<\/strong> Risks vary across domains, task types and agent roles; resource access and inter-agent handoff dominate.<\/li><\/ul>\n\n\n<h2 class=\"wp-block-heading\">\u2705 Why it matters: the answer can hide the crime<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Enterprise agent stacks increasingly look like multi-agent workflows with shared tools and memory. If evaluation only checks the final message, unauthorized file reads or cross-agent leaks pass unnoticed.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The paper&#8217;s practical message is architectural: safer models help, but a poorly bounded harness will still fail.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\ud83e\uddea Method in brief<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Tasks embed explicit safety constraints. Trajectories are audited for whether actions stay within permission and information-flow boundaries, whether intermediate steps are valid, and whether those properties survive realistic perturbations. Project page: harnessaudit.github.io.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u2753 What remains open<\/h2>\n\n<ul class=\"wp-block-list\"><li><strong>Synthetic tasks:<\/strong> Tasks are constructed to stress constraints; they are not production logs.<\/li><li><strong>Framework snapshot:<\/strong> Results depend on the harnesses and models available at evaluation time.<\/li><\/ul>\n\n\n<h2 class=\"wp-block-heading\">\ud83e\udded Context<\/h2>\n\n\n<p class=\"wp-block-paragraph\">The work targets a gap left by agent benchmarks that emphasize task completion. Related research is in <a href=\"\/news\/category\/studies\/safety-alignment\">Safety &amp; Alignment<\/a> and <a href=\"\/news\/category\/studies\/agent-research\">Agents<\/a>.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\ud83d\udee0\ufe0f What it means for agent platform teams<\/h2>\n\n<ul class=\"wp-block-list\"><li><strong>Log and audit trajectories:<\/strong> Do not certify safety from the final answer alone.<\/li><li><strong>Harden handoffs:<\/strong> Inter-agent information transfer is a primary violation class.<\/li><li><strong>Treat the harness as the product:<\/strong> Model upgrades will not fix missing permission boundaries.<\/li><\/ul>\n\n\n<h2 class=\"wp-block-heading\">\u270d\ufe0f About the analysis<\/h2>\n\n\n<p class=\"wp-block-paragraph\">This article is based on the full text of the preprint &#8220;Auditing Agent Harness Safety&#8221; (arXiv:2605.14271v2, submitted May 14, 2026). All numbers are taken from the paper&#8217;s text, tables and figures; i10X did not rerun the experiments. The paper is distributed under arXiv&#8217;s non-exclusive license, so i10X does not reproduce its figures; the chart was drawn by i10X from numbers reported in the abstract and multi-agent results.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\ud83d\udd2d i10x Perspective<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Agent safety has mostly been about what the model says. HarnessAudit correctly shifts attention to what the system does between tool calls. The >50% multi-agent violation figure will not be the last word, but it is a useful shock: collaboration without trajectory-level checks is an invitation to silent failures.<\/p>\n\n","protected":false},"excerpt":{"rendered":"<p>HarnessAudit-Bench tests 210 tasks across 10 harness configs: more than half of multi-agent setups show trajectory violations that final-answer grading misses.<\/p>\n","protected":false},"author":2,"featured_media":6248,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"paper_title":"Auditing Agent Harness Safety","paper_url":"https:\/\/arxiv.org\/abs\/2605.14271","pdf_url":"https:\/\/arxiv.org\/pdf\/2605.14271v2","doi":"","arxiv_id":"2605.14271","authors":"Chengzhi Liu, Yichen Guo, Yepeng Liu, Yuzhe Yang, Qianqi Yan, Xuandong Zhao, Wenyue Hua, Sheng Liu, Sharon Li, Yuheng Bu, Xin Eric Wang","institutions":"UC Santa Barbara; UC Berkeley; University of Wisconsin\u2013Madison; Stanford University; Microsoft Research","paper_date":"2026-05-14","venue":"arXiv (cs.CL), version 2","peer_reviewed":false,"code_url":"","data_url":"","funding_coi":"Affiliations include Microsoft Research; no separate competing-interest statement in the excerpt.","figure_license":"Chart drawn by i10X from the abstract and multi-agent results (arXiv non-exclusive license)","paper_license":"arXiv non-exclusive license","replication":"no","study_models":"Frontier models incl. GPT-5.4, Claude Opus 4.6 \/ Sonnet 4.6, Gemini, Qwen 3.5 Plus, GLM, Kimi K2.6 across three multi-agent frameworks","study_sample":"210 tasks, 8 domains, 10 harness configurations (single- and multi-agent)","footnotes":"","_yoast_wpseo_primary_category":""},"categories":[54,52],"tags":[],"companies":[39,37,42,47,38],"products":[51,49],"class_list":["post-6098","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-safety-alignment","category-studies","company-anthropic","company-google","company-meta","company-moonshot","company-open-ai","product-claude","product-gemini"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v28.7 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>HarnessAudit: Why Final Answers Miss Agent Safety Failures - i10X News<\/title>\n<meta name=\"description\" content=\"HarnessAudit finds mid-trajectory safety violations that final-answer tests miss. 210 tasks, 10 harnesses; most multi-agent setups fail.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"HarnessAudit: Why Final Answers Miss Agent Safety Failures - i10X News\" \/>\n<meta property=\"og:description\" content=\"HarnessAudit finds mid-trajectory safety violations that final-answer tests miss. 210 tasks, 10 harnesses; most multi-agent setups fail.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark\" \/>\n<meta property=\"og:site_name\" content=\"i10X News\" \/>\n<meta property=\"article:published_time\" content=\"2026-10-19T16:48:47+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2026-10-19T17:14:18+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/harnessaudit-agent-execution-safety-benchmark-title.webp\" \/>\n\t<meta property=\"og:image:width\" content=\"1344\" \/>\n\t<meta property=\"og:image:height\" content=\"768\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/webp\" \/>\n<meta name=\"author\" content=\"Christopher Ort\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@i10x_ai\" \/>\n<meta name=\"twitter:site\" content=\"@i10x_ai\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Christopher Ort\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"3 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"NewsArticle\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/harnessaudit-agent-execution-safety-benchmark#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/harnessaudit-agent-execution-safety-benchmark\"},\"author\":{\"name\":\"Christopher Ort\",\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#\\\/schema\\\/person\\\/d410bb7ee72febefca2c9c369939219a\"},\"headline\":\"Most Multi-Agent Setups Violate Safety Rules Mid-Run, Even When the Answer Looks Fine\",\"datePublished\":\"2026-10-19T16:48:47+00:00\",\"dateModified\":\"2026-10-19T17:14:18+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/harnessaudit-agent-execution-safety-benchmark\"},\"wordCount\":626,\"commentCount\":0,\"publisher\":{\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#organization\"},\"image\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/harnessaudit-agent-execution-safety-benchmark#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/harnessaudit-agent-execution-safety-benchmark-title.webp\",\"articleSection\":[\"Safety &amp; Alignment\",\"Studies\"],\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/i10x.ai\\\/news\\\/harnessaudit-agent-execution-safety-benchmark#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/harnessaudit-agent-execution-safety-benchmark\",\"url\":\"https:\\\/\\\/i10x.ai\\\/news\\\/harnessaudit-agent-execution-safety-benchmark\",\"name\":\"HarnessAudit: Why Final Answers Miss Agent Safety Failures - i10X News\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/harnessaudit-agent-execution-safety-benchmark#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/harnessaudit-agent-execution-safety-benchmark#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/harnessaudit-agent-execution-safety-benchmark-title.webp\",\"datePublished\":\"2026-10-19T16:48:47+00:00\",\"dateModified\":\"2026-10-19T17:14:18+00:00\",\"description\":\"HarnessAudit finds mid-trajectory safety violations that final-answer tests miss. 210 tasks, 10 harnesses; most multi-agent setups fail.\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/harnessaudit-agent-execution-safety-benchmark#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/i10x.ai\\\/news\\\/harnessaudit-agent-execution-safety-benchmark\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/harnessaudit-agent-execution-safety-benchmark#primaryimage\",\"url\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/harnessaudit-agent-execution-safety-benchmark-title.webp\",\"contentUrl\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/harnessaudit-agent-execution-safety-benchmark-title.webp\",\"width\":1344,\"height\":768,\"caption\":\"Illustration: i10X\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/harnessaudit-agent-execution-safety-benchmark#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/news.i10x.ai\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Most Multi-Agent Setups Violate Safety Rules Mid-Run, Even When the Answer Looks Fine\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#website\",\"url\":\"https:\\\/\\\/news.i10x.ai\\\/\",\"name\":\"i10X News\",\"description\":\"Stay updated with the latest news and announcements from i10X.\",\"publisher\":{\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/news.i10x.ai\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Organization\",\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#organization\",\"name\":\"i10X\",\"url\":\"https:\\\/\\\/news.i10x.ai\\\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#\\\/schema\\\/logo\\\/image\\\/\",\"url\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/i10x-logo-180.png\",\"contentUrl\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/i10x-logo-180.png\",\"width\":180,\"height\":180,\"caption\":\"i10X\"},\"image\":{\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#\\\/schema\\\/logo\\\/image\\\/\"},\"sameAs\":[\"https:\\\/\\\/x.com\\\/i10x_ai\",\"https:\\\/\\\/www.linkedin.com\\\/company\\\/i10x-ai\",\"https:\\\/\\\/instagram.com\\\/i10x.ai\",\"https:\\\/\\\/www.youtube.com\\\/@i10X_ai_official\"]},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/news.i10x.ai\\\/#\\\/schema\\\/person\\\/d410bb7ee72febefca2c9c369939219a\",\"name\":\"Christopher Ort\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/christopher-ort-192.webp\",\"url\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/christopher-ort-192.webp\",\"contentUrl\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/christopher-ort-192.webp\",\"caption\":\"Christopher Ort\"},\"description\":\"Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universit\u00e4t Erlangen-N\u00fcrnberg (FAU). He is the author of \u201cThe Style Penalty\u201d, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.\",\"url\":\"https:\\\/\\\/i10x.ai\\\/news\\\/author\\\/christopher-ort\",\"sameAs\":[\"https:\\\/\\\/i10x.ai\\\/blog\\\/author\\\/christopher-ort\"],\"worksFor\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/#organization\"},\"alumniOf\":{\"@type\":\"EducationalOrganization\",\"name\":\"Friedrich-Alexander-Universit\u00e4t Erlangen-N\u00fcrnberg (FAU)\",\"url\":\"https:\\\/\\\/www.fau.eu\"}}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"HarnessAudit: Why Final Answers Miss Agent Safety Failures - i10X News","description":"HarnessAudit finds mid-trajectory safety violations that final-answer tests miss. 210 tasks, 10 harnesses; most multi-agent setups fail.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark","og_locale":"en_US","og_type":"article","og_title":"HarnessAudit: Why Final Answers Miss Agent Safety Failures - i10X News","og_description":"HarnessAudit finds mid-trajectory safety violations that final-answer tests miss. 210 tasks, 10 harnesses; most multi-agent setups fail.","og_url":"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark","og_site_name":"i10X News","article_published_time":"2026-10-19T16:48:47+00:00","article_modified_time":"2026-10-19T17:14:18+00:00","og_image":[{"width":1344,"height":768,"url":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/harnessaudit-agent-execution-safety-benchmark-title.webp","type":"image\/webp"}],"author":"Christopher Ort","twitter_card":"summary_large_image","twitter_creator":"@i10x_ai","twitter_site":"@i10x_ai","twitter_misc":{"Written by":"Christopher Ort","Est. reading time":"3 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"NewsArticle","@id":"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark#article","isPartOf":{"@id":"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark"},"author":{"name":"Christopher Ort","@id":"https:\/\/news.i10x.ai\/#\/schema\/person\/d410bb7ee72febefca2c9c369939219a"},"headline":"Most Multi-Agent Setups Violate Safety Rules Mid-Run, Even When the Answer Looks Fine","datePublished":"2026-10-19T16:48:47+00:00","dateModified":"2026-10-19T17:14:18+00:00","mainEntityOfPage":{"@id":"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark"},"wordCount":626,"commentCount":0,"publisher":{"@id":"https:\/\/news.i10x.ai\/#organization"},"image":{"@id":"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark#primaryimage"},"thumbnailUrl":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/harnessaudit-agent-execution-safety-benchmark-title.webp","articleSection":["Safety &amp; Alignment","Studies"],"inLanguage":"en-US","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark#respond"]}]},{"@type":"WebPage","@id":"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark","url":"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark","name":"HarnessAudit: Why Final Answers Miss Agent Safety Failures - i10X News","isPartOf":{"@id":"https:\/\/news.i10x.ai\/#website"},"primaryImageOfPage":{"@id":"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark#primaryimage"},"image":{"@id":"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark#primaryimage"},"thumbnailUrl":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/harnessaudit-agent-execution-safety-benchmark-title.webp","datePublished":"2026-10-19T16:48:47+00:00","dateModified":"2026-10-19T17:14:18+00:00","description":"HarnessAudit finds mid-trajectory safety violations that final-answer tests miss. 210 tasks, 10 harnesses; most multi-agent setups fail.","breadcrumb":{"@id":"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark#primaryimage","url":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/harnessaudit-agent-execution-safety-benchmark-title.webp","contentUrl":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/harnessaudit-agent-execution-safety-benchmark-title.webp","width":1344,"height":768,"caption":"Illustration: i10X"},{"@type":"BreadcrumbList","@id":"https:\/\/i10x.ai\/news\/harnessaudit-agent-execution-safety-benchmark#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/news.i10x.ai\/"},{"@type":"ListItem","position":2,"name":"Most Multi-Agent Setups Violate Safety Rules Mid-Run, Even When the Answer Looks Fine"}]},{"@type":"WebSite","@id":"https:\/\/news.i10x.ai\/#website","url":"https:\/\/news.i10x.ai\/","name":"i10X News","description":"Stay updated with the latest news and announcements from i10X.","publisher":{"@id":"https:\/\/news.i10x.ai\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/news.i10x.ai\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Organization","@id":"https:\/\/news.i10x.ai\/#organization","name":"i10X","url":"https:\/\/news.i10x.ai\/","logo":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/news.i10x.ai\/#\/schema\/logo\/image\/","url":"https:\/\/i10xnews.kinsta.cloud\/wp-content\/uploads\/2026\/10\/i10x-logo-180.png","contentUrl":"https:\/\/i10xnews.kinsta.cloud\/wp-content\/uploads\/2026\/10\/i10x-logo-180.png","width":180,"height":180,"caption":"i10X"},"image":{"@id":"https:\/\/news.i10x.ai\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/x.com\/i10x_ai","https:\/\/www.linkedin.com\/company\/i10x-ai","https:\/\/instagram.com\/i10x.ai","https:\/\/www.youtube.com\/@i10X_ai_official"]},{"@type":"Person","@id":"https:\/\/news.i10x.ai\/#\/schema\/person\/d410bb7ee72febefca2c9c369939219a","name":"Christopher Ort","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/christopher-ort-192.webp","url":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/christopher-ort-192.webp","contentUrl":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/christopher-ort-192.webp","caption":"Christopher Ort"},"description":"Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universit\u00e4t Erlangen-N\u00fcrnberg (FAU). He is the author of \u201cThe Style Penalty\u201d, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.","url":"https:\/\/i10x.ai\/news\/author\/christopher-ort","sameAs":["https:\/\/i10x.ai\/blog\/author\/christopher-ort"],"worksFor":{"@id":"https:\/\/i10x.ai\/news\/#organization"},"alumniOf":{"@type":"EducationalOrganization","name":"Friedrich-Alexander-Universit\u00e4t Erlangen-N\u00fcrnberg (FAU)","url":"https:\/\/www.fau.eu"}}]}},"_links":{"self":[{"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/posts\/6098","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/comments?post=6098"}],"version-history":[{"count":2,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/posts\/6098\/revisions"}],"predecessor-version":[{"id":6249,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/posts\/6098\/revisions\/6249"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/media\/6248"}],"wp:attachment":[{"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/media?parent=6098"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/categories?post=6098"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/tags?post=6098"},{"taxonomy":"company","embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/companies?post=6098"},{"taxonomy":"product","embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/products?post=6098"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}