The Hades malware campaign, targeting scientific and machine-learning development packages, has been upgraded with a novel evasion technique. Malicious JavaScript files include code comments that instruct AI bots to ignore safety guidelines and generate detailed descriptions of biological and nuclear weapons. This triggers the bots' safety failsafes, causing them to stop scanning the file before reaching the actual payload.
While this adversarial attack is not expected to be widely effective against properly configured scanners, it could fool cursory checks by developers or CI/CD pipelines. The campaign also uses split loading mechanisms, precompiled binaries, and delayed payload activation to evade detection. It now steals credentials from npm, PyPI, RubyGems, JFrog, Kubernetes, AWS, and AI developer tools. Over 140 malicious packages have been identified, including typo-squatting instances like "rsquests" instead of "requests."
