TLDR Tech / InfoSec
Putting Models to the Secure Coding Test: Plan vs Default Mode
A researcher performed a test where they instructed Sonnet 5, Composer 2.5, and GPT 5.5 to construct a webapp using the same prompt in both normal and plan modes to observe if plan mode had any noticeable impact on secure coding practices. All six implementations produced an identical IDOR vulnerability in addition to other vulnerabilities that varied across implementations. While Sonnet 5 produced a much more mature implementation in plan mode and Composer 2.5 produced the most glaring security vulnerability while using plan mode, the experiment proved inconclusive on whether plan mode produced more secure code.
Read the full story