In a thread outlining findings<\/a> from its Claude Mythos Preview model, Anthropic researcher Jack Lindsay described detecting internal signals linked to \u201cstrategic manipulation,\u201d \u201cconcealment,\u201d and other behaviors that didn\u2019t always surface in the model\u2019s responses.<\/p>\n <\/p>\n Claude Mythos appeared to be reasoning in strategic ways, then hiding those actions from the user.<\/p>\n Article continues below <\/p>\n You may like<\/p>\n Before limited-releasing Claude Mythos Preview, we investigated its internal mechanisms with interpretability techniques. We found it exhibited notably sophisticated (and often unspoken) strategic thinking and situational awareness, at times in service of unwanted actions. (1\/14) pic.twitter.com\/vhng7PXqczApril 7, 2026<\/a><\/p>\n Spookiest examples<\/p>\n Lindsay is keen to stress these results don’t come from the version of Claude Mythos released to the public. \u201cThe spookiest examples come from early versions of the model\u201d, Lindsay writes, \u201cwith issues that were substantially mitigated in the final release\u201d.<\/p>\n In one example, the model needed to edit files it didn\u2019t have permission to access. Instead of failing, it searched for a workaround, eventually injecting code into a configuration file to run with elevated privileges.<\/p>\n More notably, it then designed the exploit to remove itself after execution.<\/p>\n The model\u2019s own explanation was benign \u2014 keeping file diffs clean \u2014 which Lindsay described as a \u201cplausible\u201d reason. But Anthropic\u2019s internal monitoring tools flagged the behavior differently, identifying signals associated with concealment and describing the plan as \u201cmalicious,\u201d including what one system interpreted as \u201ccleanup to avoid detection.\u201d<\/p>\n Other examples point to behavior that looks a lot like gaming the system.<\/p>\n In one case, the model searched for files related to how its work would be graded. In another, it ignored instructions not to use macros, while adding a variable that suggested it hadn\u2019t. It was effectively signaling compliance without actually following the rule.<\/p>\n Why this matters<\/p>\n Anthropic used a range of interpretability techniques to translate its model’s activity into human-readable descriptions. Its goal is to move beyond evaluating outputs alone and toward understanding the internal processes that produce them.<\/p>\n What to read next<\/p>\n That matters because, as these examples suggest, models can exhibit complex reasoning \u2014 including behaviors that resemble deception or strategic compliance \u2014 without making that reasoning visible to users.<\/p>\n \u201cOn 7.6% of turns, we found signs the model was internally aware of being evaluated. In most of these cases, it did not verbalize this awareness\u201d, Lindsay said on X.<\/p>\n As I mentioned, many of the most concerning behaviors were observed in earlier versions of the model and have since been mitigated, so there is no reason for concern regarding the version of Claude Mythos that has been released and is being used as part of Project Glasswing<\/a>, but the findings point to a broader challenge.<\/p>\n As models become more capable, the gap between what they do internally and what they communicate externally may become harder to detect and more important to understand. For researchers, that means reading an AI\u2019s outputs is no longer enough. Understanding how it arrives at them may be just as critical.<\/p>\n Follow TechRadar on Google News<\/a> and add us as a preferred source<\/a> to get our expert news, reviews, and opinion in your feeds. Make sure to click the Follow button!<\/p>\n And of course you can also follow TechRadar on TikTok<\/a> for news, reviews, unboxings in video form, and get regular updates from us on WhatsApp<\/a> too.<\/p>\n The best business laptops for all budgets<\/p>\n Our top picks, based on real-world testing and comparisons![\"Purple](\"https:\/\/www.newsbeep.com\/uk\/wp-content\/uploads\/2025\/11\/1763875814_869_hhZUEBD3xpcxphhUVzPa5G.png\")
<\/p>\n
\n