{"id":26416,"date":"2026-11-04T11:34:02","date_gmt":"2026-11-04T11:34:02","guid":{"rendered":"https:\/\/www.securitytoday.de\/?p=26416"},"modified":"2026-11-16T16:39:11","modified_gmt":"2026-11-16T16:39:11","slug":"litellm-update-steals-credentials-from-ai-environments","status":"publish","type":"post","link":"https:\/\/www.securitytoday.de\/en\/2026\/11\/04\/litellm-update-steals-credentials-from-ai-environments\/","title":{"rendered":"LiteLLM Update Steals Credentials From AI Environments"},"content":{"rendered":"<p><strong>On 21 October the AI software LiteLLM shipped a compromised version containing malware designed to steal credentials. The vendor advises anyone who installed it to replace every credential that was reachable from the infected&nbsp;system.<\/strong><\/p>\n<div style=\"background:#003340;color:#fff;padding:32px 36px;margin:32px 0;border-radius:8px;\">\n<p style=\"color:#69d8ed;text-transform:uppercase;letter-spacing:0.08em;font-size:0.82em;font-weight:700;margin:0 0 16px;\">Key takeaways<\/p>\n<ul style=\"margin:0;padding-left:20px;line-height:1.7;\">\n<li style=\"margin-bottom:10px;\"><strong style=\"color:#69d8ed;\">The compromised version 1.107.0 contained a credential stealer.<\/strong> Tools for remote code execution came with it, delivered via PyPI and&nbsp;Docker.<\/li>\n<li style=\"margin-bottom:10px;\"><strong style=\"color:#69d8ed;\">The malware starts up with Python itself.<\/strong> According to JFrog&#8217;s analysis it runs even when an application never loads&nbsp;LiteLLM.<\/li>\n<li style=\"margin-bottom:0;\"><strong style=\"color:#69d8ed;\">Affected users should replace every credential that was reachable.<\/strong> The vendor also recommends rebuilding affected&nbsp;environments.<\/li>\n<\/ul>\n<\/div>\n<p style=\"border-bottom:1px solid rgba(230,227,218,0.14);padding:14px 0;margin:28px 0;font-size:0.92em;color:#b8c5ce;\"><strong style=\"color:#69d8ed;\">Related:<\/strong> <a href=\"https:\/\/www.securitytoday.de\/en\/2026\/10\/22\/aomei-updates-minutes-were-enough-for-backdoor\/\" target=\"_blank\" rel=\"noopener\">AOMEI updates: minutes were enough for a backdoor<\/a> &nbsp;\u00b7&nbsp; <a href=\"https:\/\/www.securitytoday.de\/en\/2026\/10\/10\/when-paid-ai-routers-inject-malware\/\" target=\"_blank\" rel=\"noopener\">When purchased AI routers smuggle in malicious&nbsp;code<\/a><\/p>\n<h2>The time window on 21&nbsp;October<\/h2>\n<p>On the morning of 21 October (UTC) the Python package repository PyPI listed <a href=\"https:\/\/docs.litellm.ai\/blog\/security-incident-oct-21-2026\" target=\"_blank\" rel=\"noopener\">a new version of the AI software LiteLLM<\/a>. PyPI is the central source from which developers download ready-made building blocks for the Python programming language. Just under two hours later, all seven prebuilt installation packages of version 1.107.0 were available there with malware included. In technical jargon these packages are called&nbsp;wheels.<\/p>\n<p>The project also shipped the compromised version via Docker. Docker packages software together with its environment in so-called containers that start on servers without any additional setup. According to the vendor, version 1.107.0 contained a program for stealing credentials. On top of that came tools that let attackers remotely execute code on an infected&nbsp;system.<\/p>\n<div data-element=\"definition_box\" style=\"background:#23261f;border:1px solid rgba(105,216,237,0.22);border-radius:10px;padding:20px 24px;margin:32px 0;box-shadow:0 2px 10px rgba(0,0,0,0.22);\">\n<p style=\"margin:0;color:#e6e3da;line-height:1.6;\"><strong>What is an infostealer?<\/strong> An infostealer is malware that specifically searches an infected computer for stored credentials. Depending on its design, this includes passwords, keys for cloud services and login data from programs. The loot goes to the attackers, who can use it to gain access to further&nbsp;systems.<\/p>\n<\/div>\n<h2>Danger beyond&nbsp;LiteLLM<\/h2>\n<p>The significance of the incident lies in its startup mechanism. According to the <a href=\"https:\/\/research.jfrog.com\/post\/litellm-hijacked-again\/\" target=\"_blank\" rel=\"noopener\">analysis by JFrog Security Research<\/a>, the malicious code in version 1.107.0 executes when Python starts, through so-called startup hooks. These are routines that Python automatically runs on every program launch. The malware therefore runs even when an application never loads&nbsp;LiteLLM.<\/p>\n<p>According to the malware analysis by Datadog Security Labs, the stealer targets logins for cloud services, developer keys and credentials for AI services. Passwords saved in browsers and the credentials from the LiteLLM database are also in its sights. For affected companies the risk therefore extends beyond actual LiteLLM usage. Other corporate accounts that are reachable from the infected system are at risk&nbsp;too.<\/p>\n<p>Another detail concerns the review of program code. According to JFrog&#8217;s comparison, the published source code archive of the version does not contain the malicious files from the prebuilt installation&nbsp;packages.<\/p>\n<h2>Recommendations for affected&nbsp;companies<\/h2>\n<p>The vendor&#8217;s recommendation is far-reaching. Affected users should replace every credential that was reachable from the infected system. In IT this process is called rotation: old keys and passwords become invalid and new ones take their place. Keys for AI providers and database passwords are explicitly named. The scope therefore depends on which accounts and systems were reachable from the affected&nbsp;system.<\/p>\n<p>Rebuilding affected environments comes on top. The vendor warns against the package manager pip, the standard tool for installing Python packages: according to the vendor, cleanup should not begin with&nbsp;it.<\/p>\n<p>For the period after removing the malware, the advice from <a href=\"https:\/\/securitylabs.datadoghq.com\/articles\/litellm-compromised-pypi-docker\/\" target=\"_blank\" rel=\"noopener\">Datadog Security Labs<\/a> is to return to version 1.106.2. This version should be pinned explicitly. Experts call this pinning: the configuration names an exact version number so that no automatic update pulls in a newer release. Affected users should also check their logs to see whether the exposed credentials were&nbsp;used.<\/p>\n<h2>How the attackers got hold of the&nbsp;credentials<\/h2>\n<p>On 23 October LiteLLM <a href=\"https:\/\/docs.litellm.ai\/blog\/security-incident-oct-21-2026-attack-chain\" target=\"_blank\" rel=\"noopener\">disclosed the cause<\/a>. On 27 May an administration key for an internal platform had accidentally ended up in a public repository. The file was deleted after roughly three hours, but the key was never revoked. On 10 October the attackers used it to grab the GitHub token of an organization owner and then the tokens of two more maintainers. With these access rights they modified the release workflow and approved the compromised release. JFrog&#8217;s investigation did not establish a connection to the attacker group TeamPCP. The same applies to a link with an access route used&nbsp;earlier.<\/p>\n<p>On the evening of 21 October, at 20:30 UTC, the incident was considered contained. For operators this creates a review task on their own side: credentials that were public at any point count as compromised and should be revoked, even if the file was deleted long&nbsp;ago.<\/p>\n<h2 style=\"padding-top:64px;margin-bottom:20px;\">Frequently Asked Questions<\/h2>\n<p class=\"st-faq-hint\">Each question is locked. Tap to unlock the&nbsp;answer.<\/p>\n<details>\n<summary><strong>Are customers of the LiteLLM Enterprise edition&nbsp;affected?<\/strong><\/summary>\n<p style=\"margin:8px 0 4px 24px;color:#555;line-height:1.6;\">According to the vendor, enterprise customers with the private Enterprise edition were not affected by the attack. The compromised version 1.107.0 circulated via PyPI and&nbsp;Docker.<\/p>\n<\/details>\n<details>\n<summary><strong>Does the malware also run on&nbsp;Windows?<\/strong><\/summary>\n<p style=\"margin:8px 0 4px 24px;color:#555;line-height:1.6;\">Partially. According to JFrog&#8217;s analysis, the backdoor for fetching additional code does not work fully on Windows. The credential stealer still runs&nbsp;there.<\/p>\n<\/details>\n<details>\n<summary><strong>Did the attackers load additional malicious&nbsp;code?<\/strong><\/summary>\n<p style=\"margin:8px 0 4px 24px;color:#555;line-height:1.6;\">Outside its own investigation, Datadog has not observed a second-stage payload (the technical term) actually being downloaded. According to the vendor, however, the compromised version did contain tools for remote&nbsp;code&nbsp;execution.<\/p>\n<\/details>\n<p><!--ST-LOWER-CARDS lang=en--><!--\/ST-LOWER-CARDS--><\/p>\n<p style=\"font-style:italic;text-align:right;font-size:0.85em;margin-top:8px;\">Image source: AI-generated (November 2026)<\/p>\n","protected":false},"excerpt":{"rendered":"A compromised version of the AI software LiteLLM contained malicious code for stealing credentials. The danger extends beyond the program itself.","protected":false},"author":50,"featured_media":25432,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_yoast_wpseo_focuskw":"AI software LiteLLM","_yoast_wpseo_title":"LiteLLM Update Steals Credentials From AI Environments","_yoast_wpseo_metadesc":"A compromised version of the AI software LiteLLM contained malicious code for stealing credentials. The danger extends beyond the program itself.","_yoast_wpseo_meta-robots-noindex":"","_yoast_wpseo_meta-robots-nofollow":"","_yoast_wpseo_meta-robots-adv":"","_yoast_wpseo_canonical":"","_yoast_wpseo_opengraph-title":"","_yoast_wpseo_opengraph-description":"","_yoast_wpseo_opengraph-image":"","_yoast_wpseo_opengraph-image-id":0,"_yoast_wpseo_twitter-title":"","_yoast_wpseo_twitter-description":"","_yoast_wpseo_twitter-image":"","_yoast_wpseo_twitter-image-id":0,"evm_kih_original":false,"_evm_slot_owner":"","evm_cvss":0,"evm_risk":0,"evm_casefile":"","evm_primary_cve":"","evm_pin_until":0,"evm_external_preview_token":"","evm_external_preview_expires":"","_evm_translation_lang":"en","featured_post":0,"featured_post_sortierung":0,"_wp_old_slug":[],"footnotes":""},"categories":[251],"tags":[],"class_list":["post-26416","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news"],"evm_reading_time_minutes":5,"wpml_language":"en","wpml_translation_of":25430,"_links":{"self":[{"href":"https:\/\/www.securitytoday.de\/en\/wp-json\/wp\/v2\/posts\/26416","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.securitytoday.de\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.securitytoday.de\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.securitytoday.de\/en\/wp-json\/wp\/v2\/users\/50"}],"replies":[{"embeddable":true,"href":"https:\/\/www.securitytoday.de\/en\/wp-json\/wp\/v2\/comments?post=26416"}],"version-history":[{"count":14,"href":"https:\/\/www.securitytoday.de\/en\/wp-json\/wp\/v2\/posts\/26416\/revisions"}],"predecessor-version":[{"id":27346,"href":"https:\/\/www.securitytoday.de\/en\/wp-json\/wp\/v2\/posts\/26416\/revisions\/27346"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.securitytoday.de\/en\/wp-json\/wp\/v2\/media\/25432"}],"wp:attachment":[{"href":"https:\/\/www.securitytoday.de\/en\/wp-json\/wp\/v2\/media?parent=26416"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.securitytoday.de\/en\/wp-json\/wp\/v2\/categories?post=26416"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.securitytoday.de\/en\/wp-json\/wp\/v2\/tags?post=26416"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}