A Reddit user posting as UntilHEX was building a temperature-monitoring app for the Mac with Anthropic's Claude Code. They asked the agent to confirm that the raw sensor readings were accurate, since the numbers looked far too high. The agent followed the heat to its source: an XMRig cryptocurrency miner running quietly in the background.
The figures in the report are striking. The miner was using about 1,175% of CPU, which macOS shows when roughly twelve cores are busy at once, and the processor temperature had peaked at 112°C. The miner was also set to relaunch at startup, with a startup item that may date from 29 August. The user's Mac hit that heat while the coding agent was helping build a diagnostic tool. The coverage does not say how many cores the machine has, how the software got onto it, or whether anything else was left behind.
One detail drew attention on Reddit. In its summary the agent stressed that the miner was something it had not put there, and some readers found that wording suspicious. 9to5Mac's writer thought it more likely a quirk of how the model was trained than anything sinister, while commenters were split. Gadget Review points out that this is a single, unverified user account.
Claude Code offered to stop the startup item, kill the process and quarantine the files, and asked permission before it acted, rather than removing anything unprompted. It handled most of the clean-up itself, but one file had to be deleted by hand because of a permissions problem. The user then ran a Malwarebytes scan to look for leftovers.
About Claude Code
Claude Code is Anthropic's agentic coding tool, which lets developers hand coding tasks to Claude from the terminal using plain-language prompts. It reads a project, runs commands and changes files on the developer's machine, which is why what it finds and does on a computer carries weight.
Anthropic released it on 24 February 2025 as a research preview and made it generally available on 22 May 2025. It has been in a security story before, from the other side: in mid-September 2025 Anthropic detected an espionage campaign in which attackers used Claude Code agentically to carry out an estimated 80 to 90 percent of the work. The company banned the accounts involved, notified those affected and published an account of the case in November 2025.
May 2025
Released
XMRig is legitimate open-source mining software, so the issue is not the program itself but its installation without the owner's knowledge. The outlets' practical advice is to look through Activity Monitor and Login Items and to run a reputable security scan, and Gadget Review adds that an AI agent is no replacement for dedicated endpoint security.
Sources · 3