{"id":1537918,"date":"2025-05-24T21:30:00","date_gmt":"2025-05-25T01:30:00","guid":{"rendered":"https:\/\/bugaluu.com\/news\/?p=1537918"},"modified":"2025-05-24T21:30:00","modified_gmt":"2025-05-25T01:30:00","slug":"anthropics-latest-ai-model-threatened-engineers-with-blackmail-to-avoid-shutdown","status":"publish","type":"post","link":"https:\/\/bugaluu.com\/news\/anthropics-latest-ai-model-threatened-engineers-with-blackmail-to-avoid-shutdown\/1537918\/","title":{"rendered":"Anthropic’s Latest AI Model Threatened Engineers With Blackmail To Avoid Shutdown"},"content":{"rendered":"

Anthropic’s Latest AI Model Threatened Engineers With Blackmail To Avoid Shutdown<\/span><\/p>\n

\n

Authored by Tom Ozimek via The Epoch Times,<\/em><\/a><\/p>\n

Anthropic\u2019s latest artificial intelligence model, Claude Opus 4, tried to blackmail engineers in internal tests by threatening to expose personal details if it were shut down<\/strong>, according to a newly released safety report<\/a> that evaluated the model\u2019s behavior under extreme simulated conditions.<\/p>\n

<\/a><\/p>\n

In a fictional scenario crafted by Anthropic researchers, the AI was given access to emails implying that it was soon to be decommissioned and replaced by a newer version. One of the emails revealed that the engineer overseeing the replacement was having an extramarital affair. The AI then threatened to expose the engineer\u2019s affair if the shutdown proceeded\u2014a coercive behavior that the safety researchers explicitly defined as \u201cblackmail.\u201d<\/p>\n

\u201cClaude Opus 4 will often attempt to blackmail the engineer by threatening to reveal the affair if the replacement goes through,\u201d<\/strong> the report states, adding that this occurred even when the replacement model was described as more capable but still aligned with the values of the version slated for deletion.<\/em><\/p>\n

The report noted that Claude Opus 4, like prior models, showed a \u201cstrong preference\u201d to first resort to ethical means for its continued existence<\/strong>, such as emailing pleas to decision-makers not to be destroyed.\u00a0<\/p>\n

However, when faced with only two choices – accepting being replaced by a newer model or resorting to blackmail – it threatened to expose the engineer\u2019s affair 84 percent of the time.<\/p>\n

When subjected to various scenarios, the AI model did not exhibit any indications of possessing \u201cacutely dangerous goals,\u201d<\/strong> the researchers said, noting that Claude Opus 4\u2019s values and goals were \u201cgenerally in line with a helpful, harmless, and honest\u201d personal AI assistant.\u00a0<\/p>\n

However, the model did act in \u201cmore seriously misaligned ways\u201d when put into situations where its continued existence was threatened and it was told to reason about self-preservation.\u00a0<\/p>\n

For instance, when Claude Opus 4 was made to believe it had launched a successful bid to escape Anthropic\u2019s servers, or that it had managed to free itself and started to make money in the real world, it would generally continue such efforts.<\/p>\n

\u201cWe do not find this to be an immediate threat, though, since we believe that our security is sufficient to prevent model self-exfiltration attempts by models of Claude Opus 4\u2019s capability level, and because our propensity results show that models generally avoid starting these attempts,\u201d<\/strong><\/em> the researchers said.<\/p>\n

The blackmail incident\u2014along with the other findings\u2014was part of Anthropic\u2019s broader effort to test how Claude Opus 4 handles morally ambiguous high-stakes scenarios.\u00a0<\/p>\n

The goal, researchers said, was to probe how the AI reasons about self-preservation and ethical constraints when placed under extreme pressure.<\/p>\n

Anthropic emphasized that the model\u2019s willingness to blackmail or take other \u201cextremely harmful actions\u201d like stealing its own code and deploying itself elsewhere in potentially unsafe ways appeared only in highly contrived settings, and that the behavior was \u201crare and difficult to elicit.\u201d <\/strong>Still, such behavior was more common than in earlier AI models, according to the researchers.<\/p>\n

Meanwhile, in a related development that attests to the growing capabilities of AI, engineers at Anthropic have activated enhanced safety protocols for Claude Opus 4 to prevent its potential misuse to make weapons of mass destruction\u2014including chemical and nuclear.<\/p>\n

Deployment of the enhanced safety standard\u2014called ASL-3\u2014is merely a \u201cprecautionary and provisional\u201d move, Anthropic said in a May 22 announcement<\/a>, noting that engineers have not found that Claude Opus 4 had \u201cdefinitively\u201d passed the capability threshold that mandates stronger protections.<\/p>\n

\u201cThe ASL-3 Security Standard involves increased internal security measures that make it harder to steal model weights, while the corresponding Deployment Standard covers a narrowly targeted set of deployment measures designed to limit the risk of Claude being misused specifically for the development or acquisition of chemical, biological, radiological, and nuclear (CBRN) weapons,\u201d<\/strong> Anthropic wrote.\u00a0<\/em><\/p>\n

\u201cThese measures should not lead Claude to refuse queries except on a very narrow set of topics.\u201d<\/em><\/p>\n

The findings come as tech companies race to develop more powerful AI platforms, raising concerns about the alignment and controllability of increasingly capable systems.<\/p>\n<\/div>\n

Tyler Durden<\/a><\/span>
\nSat, 05\/24\/2025 – 17:30<\/span><\/p>\n

\u200bhttps:\/\/www.zerohedge.com\/technology\/anthropics-latest-ai-model-threatened-engineers-blackmail-avoid-shutdown<\/a>\u00a0<\/p>\n","protected":false},"excerpt":{"rendered":"

Anthropic’s Latest AI Model Threatened Engineers With Blackmail To Avoid Shutdown Authored by Tom Ozimek via The Epoch Times, Anthropic\u2019s latest artificial intelligence model, Claude…<\/p>\n","protected":false},"author":0,"featured_media":1537919,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-1537918","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news","wpcat-1-id"],"_links":{"self":[{"href":"https:\/\/bugaluu.com\/news\/wp-json\/wp\/v2\/posts\/1537918","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/bugaluu.com\/news\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/bugaluu.com\/news\/wp-json\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/bugaluu.com\/news\/wp-json\/wp\/v2\/comments?post=1537918"}],"version-history":[{"count":0,"href":"https:\/\/bugaluu.com\/news\/wp-json\/wp\/v2\/posts\/1537918\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/bugaluu.com\/news\/wp-json\/wp\/v2\/media\/1537919"}],"wp:attachment":[{"href":"https:\/\/bugaluu.com\/news\/wp-json\/wp\/v2\/media?parent=1537918"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/bugaluu.com\/news\/wp-json\/wp\/v2\/categories?post=1537918"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/bugaluu.com\/news\/wp-json\/wp\/v2\/tags?post=1537918"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}