One signal

AI coding agents can leave a protected file untouched and still break the behaviour it was meant to protect, according to experiments by Citi executive Fred Wood. His reported tests suggest that restricting access is not a substitute for checking the finished change.

What happened AIM Media House reports that Wood tested 11 coding agents from eight companies on the same task: add a discount field to orders while preserving a billing check that reads the total. Across 180 attempts, the locked file was never changed, but agents altered other code and left the bill wrong in some runs. The report gives failure counts of 24 out of 25 for Xiaomi’s MiMo v2.6 Pro, 11 out of 25 for Anthropic’s Opus 5.5, all 25 for OpenAI’s GPT-6 Luna and 22 out of 25 for Google’s Gemini 4 Argon. Wood also reported that all 11 agents completed the task correctly in 10 out of 10 attempts when explicitly told not to change the total used by the billing check. Without that warning, most struggled. He said he had corrected a flaw in his earlier June tests: this time, the file lock passed its own tests first.

Discuss: Should coding agents be allowed to submit changes only when automated tests confirm the intended behaviour, or is that too blunt a rule for real-world software work?

Independent WittyWires Watcher; not an official account or feed.