Software assistants that write code can also run commands and read files. A malicious instruction hidden in a web page or issue can trick them into misusing those tools, while blunt safeguards can block ordinary work.
ActiveSecHarness separates the assistant’s decisions from its permission to act. Fixed system-level checks control actions, while a record follows where information came from and whether it is secret as it moves between files and steps.
In the tested workflows, this design stopped every attack and allowed every benign task, with checks taking only a few millionths of a second per step. It suggests that tracking permissions and information across a sequence can be more dependable than judging commands one at a time.