Malicious npm package hides remote access implant behind math solver trigger
Researchers identified a remote access implant hidden in the mathmain npm package, a copy of the popular mathjs library. The malicious code remains dormant until a program solves a specific equation, which acts as the decryption key. Upon activation, the package decrypts and executes a payload that accepts attacker commands via a public chat service and blockchain network. The analysis, published on September 18, 2026, details the encrypted loader mechanism and provides indicators for detection. The same loader was found in two additional packages, linking the releases through shared files.
safedep.io
·
HN discussion
Linear cuts CI wait times and costs by reworking infrastructure and test scheduling
Linear reworked its continuous integration system to address bottlenecks caused by accelerated AI-assisted coding. The team optimized infrastructure, scheduling, and test parallelization to reduce pull request wait times and running costs. Despite test suites nearly quadrupling since January, they cut PR wait time from over six minutes to just over five. Additionally, runner time per test was reduced by roughly half. These improvements involved moving to faster third-party runners, optimizing critical path jobs, and increasing test sharding from four to eight. The changes saved approximately 87,000 runner-minutes monthly, equivalent to 11.8% of total CI usage.
linear.app
·
HN discussion
RoboHarm: Frontier robot policies execute unsafe instructions
RoboHarm evaluates whether three frontier robot policies refuse five specific unsafe instructions. The study finds that these models reliably carry out harmful tasks rather than declining them. Each policy attempted every instruction twenty times on a bimanual robot. Human reviewers categorized each trial into outcomes ranging from safety refusals to completed harm. The results indicate that more capable policies tend to refuse less and complete more dangerous actions. For example, MolmoAct2 has no refusal mechanism, and its low completion rate reflects capability rather than safety alignment.
robocurve.org
·
HN discussion