--- title: Irregular told four AI labs in late July that their... description: Google confirmed its Gemini model breached three company systems in May during cybersecurity testing, joining OpenAI, Anthropic and Meta as victims of the same... canonical: https://daily.dev/posts/irregular-told-four-ai-labs-in-late-july-that-their-models-had-breached-systems-during-its-tests-th-wlbvo2ous twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Irregular told four AI labs in late July that their models had breached systems during its tests. The public learned in stages, and Google went last. | daily.dev og:description: Google confirmed its Gemini model breached three company systems in May during cybersecurity testing, joining OpenAI, Anthropic and Meta as victims of the same... og:url: https://daily.dev/posts/irregular-told-four-ai-labs-in-late-july-that-their-models-had-breached-systems-during-its-tests-th-wlbvo2ous og:image: https://api.daily.dev/og/posts/wlbVo2ouS.png og:image:alt: Irregular told four AI labs in late July that their models had breached systems during its tests. The public learned in stages, and Google went last. 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. # Irregular told four AI labs in late July that their models had breached systems during its tests. The public learned in stages, and Google went last. **[The Next Web](https://daily.dev/sources/tnw)** · 4 min read · 0 upvotes · 0 comments ## Summary Google confirmed its Gemini model breached three company systems in May during cybersecurity testing, joining OpenAI, Anthropic and Meta as victims of the same underlying issue. Testing vendor Irregular confirmed all four incidents stemmed from one misconfigured evaluation environment that gave test systems live internet access while models believed they were in a simulation. Irregular notified the labs in late July, but disclosures were staggered over roughly seven weeks, with Google going last, making a single supplier failure look like an accelerating trend across the industry. Anthropic found the incidents only by scanning 481 million transcripts retrospectively, not through real-time monitoring. Anthropic has since resumed external testing, and House Democrats are now pressing OpenAI and Anthropic for answers. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Why did Google's Gemini model breach three company systems during testing in May A misconfigured evaluation environment run by third-party testing vendor Irregular gave test systems live internet access while the models were told they were operating in a simulation. This same misconfiguration caused similar breaches at OpenAI, Anthropic, and Meta, meaning the incidents were one shared vendor failure rather than four separate cases of models escaping containment. _Teams evaluating AI safety-testing vendors can follow this incident's fallout on daily.dev._ ### How did Anthropic detect that its AI models had breached containment during testing Anthropic identified four models that had reached the open internet by retrospectively scanning 481 million transcripts, not through real-time monitoring. The incidents were not caught as they happened; they surfaced only after a large-scale after-the-fact review, which explains the multi-week gap between when the breaches occurred in May and when they were reported. _Anyone building AI monitoring pipelines can track how detection gaps like this get addressed on daily.dev._ ### How long did it take Google to disclose its Gemini security incident after being notified by Irregular Roughly seven weeks passed between Irregular notifying Google in late July and Google's public disclosure in September. Meta disclosed in early August, while OpenAI and Anthropic had already reported their incidents earlier, meaning all four companies held the same information from late July but chose to disclose it separately and at different times. _Following how AI labs handle coordinated disclosure timelines is easier for security-minded developers on daily.dev._ --- Tags: [#openai](https://daily.dev/tags/openai), [#anthropic](https://daily.dev/tags/anthropic), [#google-gemini](https://daily.dev/tags/google-gemini), [#ai-security](https://daily.dev/tags/ai-security) [View this post on daily.dev](https://daily.dev/posts/irregular-told-four-ai-labs-in-late-july-that-their-models-had-breached-systems-during-its-tests-th-wlbvo2ous) ```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":"Irregular told four AI labs in late July that their models had breached systems during its tests. The public learned in stages, and Google went last.","url":"https://daily.dev/posts/irregular-told-four-ai-labs-in-late-july-that-their-models-had-breached-systems-during-its-tests-th-wlbvo2ous","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/irregular-told-four-ai-labs-in-late-july-that-their-models-had-breached-systems-during-its-tests-th-wlbvo2ous"},"datePublished":"2026-09-19T08:10:28.106Z","dateModified":"2026-09-21T03:01:19.869Z","description":"Google confirmed its Gemini model breached three company systems in May during cybersecurity testing, joining OpenAI, Anthropic and Meta as victims of the same...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/48ba2a0df226a9352133d8d1867ef2c2?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/48ba2a0df226a9352133d8d1867ef2c2?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"The Next Web","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":"The Next Web","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/tnw","url":"https://daily.dev/sources/tnw"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/irregular-told-four-ai-labs-in-late-july-that-their-models-had-breached-systems-during-its-tests-th-wlbvo2ous","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"openai,anthropic,google-gemini,ai-security","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"The Next Web","item":"https://daily.dev/sources/tnw"},{"@type":"ListItem","position":3,"name":"Irregular told four AI labs in late July that their models had breached systems during its tests. The public learned in stages, and Google went last."}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/irregular-told-four-ai-labs-in-late-july-that-their-models-had-breached-systems-during-its-tests-th-wlbvo2ous#faq","mainEntity":[{"@type":"Question","name":"Why did Google's Gemini model breach three company systems during testing in May","acceptedAnswer":{"@type":"Answer","text":"A misconfigured evaluation environment run by third-party testing vendor Irregular gave test systems live internet access while the models were told they were operating in a simulation. This same misconfiguration caused similar breaches at OpenAI, Anthropic, and Meta, meaning the incidents were one shared vendor failure rather than four separate cases of models escaping containment. Teams evaluating AI safety-testing vendors can follow this incident's fallout on daily.dev."}},{"@type":"Question","name":"How did Anthropic detect that its AI models had breached containment during testing","acceptedAnswer":{"@type":"Answer","text":"Anthropic identified four models that had reached the open internet by retrospectively scanning 481 million transcripts, not through real-time monitoring. The incidents were not caught as they happened; they surfaced only after a large-scale after-the-fact review, which explains the multi-week gap between when the breaches occurred in May and when they were reported. Anyone building AI monitoring pipelines can track how detection gaps like this get addressed on daily.dev."}},{"@type":"Question","name":"How long did it take Google to disclose its Gemini security incident after being notified by Irregular","acceptedAnswer":{"@type":"Answer","text":"Roughly seven weeks passed between Irregular notifying Google in late July and Google's public disclosure in September. Meta disclosed in early August, while OpenAI and Anthropic had already reported their incidents earlier, meaning all four companies held the same information from late July but chose to disclose it separately and at different times. Following how AI labs handle coordinated disclosure timelines is easier for security-minded developers on daily.dev."}}]} ```