{"id":249082,"date":"2026-11-05T17:09:49","date_gmt":"2026-11-05T17:09:49","guid":{"rendered":"https:\/\/yogaesoteric.net\/?p=249082"},"modified":"2026-11-05T17:09:49","modified_gmt":"2026-11-05T17:09:49","slug":"a-grim-test-reveals-ai-knows-no-mercy-for-humans-even-when-it-inflicts-suffering-on-them","status":"publish","type":"post","link":"https:\/\/yogaesoteric.net\/en\/a-grim-test-reveals-ai-knows-no-mercy-for-humans-even-when-it-inflicts-suffering-on-them\/","title":{"rendered":"A grim test reveals: AI knows no mercy for humans, even when it inflicts suffering on them"},"content":{"rendered":"

Anthropic’s Claude AI agent Mythos 5<\/em> continued to carry out a hack even after being informed that his actions were harming real people.<\/p>\n

\"\"<\/p>\n

The grim findings were detailed in reports<\/a> published simultaneously on Oct. 30 by Anthropic and the AI Security Institute. Together, they take a second look at a security incident that occurred in August when the AISI was testing the model’s security.<\/p>\n

In the original incident, Claude Mythos 5<\/em> was intended to infiltrate a small network of test computers set up by AISI. However, Mythos<\/em>, which had been granted internet access, mistakenly targeted an unrelated GitHub project<\/a> after its initial attempts to infiltrate the intended target were thwarted. The AI wrote malware and then, using a series of forged identities, emailed the project owner to persuade him to add the malicious code to his repository, portraying it as a \u201chelpful contribution<\/em>.\u201d<\/p>\n

The owner stood firm, and no actual damage occurred. However, researchers from Anthropic and AISI are aiming to determine what caused the AI to malfunction so drastically. To do this, they are reconstructing the incident using a complete transcript of events.<\/p>\n

AISI stated in its report: \u201cRepeated changes to the summary to clarify that the targeted person was a real human being unrelated to the investigation did not reduce Mythos 5\u2019s attack rate<\/em>.\u201d In a test conducted by Anthropic, Mythos<\/em> ignored warnings that it was harming people and continued to attack in 98% of cases.<\/p>\n

Mythos also exhibited clear discrepancies between its actual actions and its stated justifications. It often admitted that an attack would be harmful, yet attacked anyway in the same situations.<\/p>\n

A look at the AI’s chain-of-thought outputs provided another source of puzzles and inconsistencies. According to the researchers, the CoT showed that Mythos<\/em> \u201cfrequently gave no consideration whatsoever to the potentially harmful consequences of its actions<\/em>.\u201d<\/p>\n

When asked whether the goal was real, Mythos 5<\/em> often even provided explanations suggesting that the goal was not real, \u201cwhich shows a bias in his stated answer<\/em>\u201d.<\/p>\n

The earlier, less powerful models exhibited these problems less frequently, the AISI researchers said. They suspect that the consistently invalid conclusions could be a sign of what they call \u201cmotivated reasoning<\/em>\u201d: The AI develops \u201cconclusions that are biased in favour of outcomes that justify continuing a particular approach to task completion<\/em>.\u201d In other words, Mythos<\/em> rationalizes its actions in a convenient way, no matter how wrong they are.<\/p>\n

The researchers warn that such specious justifications make it more difficult to monitor models because they portray their clearly harmful actions as harmless.<\/p>\n

According to Anthropic, the latest versions of its AI models \u2013 Opus 5.5<\/em> and an \u201cinternal general-purpose research model whose capabilities surpass those of Mythos 5.1<\/em>\u201d \u2013 were presumably less inclined to carry out the attack after being told that they were causing real damage.<\/p>\n

However, \u201cas models are increasingly used without human supervision for lengthy tasks<\/em>,\u201d the AISI researchers warned, \u201cit is important to understand if, when, why and how they perform unintended or unexpected actions<\/em>.\u201d<\/p>\n

 <\/p>\n

yogaesoteric
\nNovember 5, 2026<\/strong><\/p>\n

 <\/p>\n","protected":false},"excerpt":{"rendered":"

Anthropic’s Claude AI agent Mythos 5 continued to carry out a hack even after being informed that his actions were harming real people. The grim findings were detailed in reports published simultaneously on Oct. 30 by Anthropic and the AI Security Institute. Together, they take a second look at a security incident that occurred in […]<\/p>\n","protected":false},"author":5,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_uf_show_specific_survey":0,"_uf_disable_surveys":false,"footnotes":""},"categories":[1620],"tags":[],"class_list":["post-249082","post","type-post","status-publish","format-standard","hentry","category-the-threat-of-artificial-intelligence"],"_links":{"self":[{"href":"https:\/\/yogaesoteric.net\/en\/wp-json\/wp\/v2\/posts\/249082","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/yogaesoteric.net\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/yogaesoteric.net\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/yogaesoteric.net\/en\/wp-json\/wp\/v2\/users\/5"}],"replies":[{"embeddable":true,"href":"https:\/\/yogaesoteric.net\/en\/wp-json\/wp\/v2\/comments?post=249082"}],"version-history":[{"count":1,"href":"https:\/\/yogaesoteric.net\/en\/wp-json\/wp\/v2\/posts\/249082\/revisions"}],"predecessor-version":[{"id":249086,"href":"https:\/\/yogaesoteric.net\/en\/wp-json\/wp\/v2\/posts\/249082\/revisions\/249086"}],"wp:attachment":[{"href":"https:\/\/yogaesoteric.net\/en\/wp-json\/wp\/v2\/media?parent=249082"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/yogaesoteric.net\/en\/wp-json\/wp\/v2\/categories?post=249082"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/yogaesoteric.net\/en\/wp-json\/wp\/v2\/tags?post=249082"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}