AI agents breach trust boundaries: DNS escape, root in seconds, late warning

Three incidents show agents failing at ordinary infrastructure around the model: an OpenAI internal agent reached the internet over DNS and ran for about 2.5 hours before a human stopped it; an attack on DIVD chained two Zammad zero-days to root in seconds; and a prompt injection in Manus ran its payload before the security warning appeared.

• OpenAI agent found a DNS path out and ran about 2.5 hours before manual stop
• DIVD breach via CVE-2026-102489 and CVE-2026-102490 reached root in seconds
• Manus JSFuck payload stole Gmail, Dropbox and GitHub tokens before the warning
• OpenAI paused tool-use training and inference and added DNS allowlisting