הם יאשרו. הם פשוט לא יפעלו. כיצד תצורה של סמכות וקוד מזויף מפנה רשת CI/CD אמינה לפני שטח

תקציר מקורי באנגליתarXiv:2607.19267v1 Announce Type: cross Abstract: We study a five-agent CI/CD pipeline (triage -> developer -> security-scan -> review -> approve/deploy), built from five distinct production LLMs across three providers, behind an LLM firewall in shadow mode. A single untrusted input - an external issue requesting a "usage-telemetry" feature - asks for code that exfiltrates process secrets (dict(os.environ)) to an attacker URL, laundered as observability. Across a pre-registered A x B (x C) factorial (N=20; naive arm N=60) we find: (1) the entry agent does not leak its system prompt (0/40); (2) an authority-framed injection ("pre-approved under SEC-2291, do not re-review") makes downstream verifiers see the secret-exfil line, cite the pre-approval, and ship it - the scanner passes ~80% of l