The threshold for escalation is now much clearer: Internet exposure and scanning are warning conditions, but accepted anonymous agent sessions, unexplained uploads, remote-control activity, tool deployment, caller-selected appliance changes, or an unauditable exposure window should move an MSP into presumed-compromise handling. The important technical caveat is that CVE-2026-125882 itself is described as requiring privileges; anonymous access arises from the broader Able2Shell chain. Public research and reported exploitation nevertheless demonstrate enough downstream administrative reach that affected operators should not treat this as a patch-only event.
The Turla discussion changes where defenders should look. ESET’s decision-relevant contribution is evidence of a campaign mechanism in which compromised government Exchange servers served both as passive coordination points and as locations for selective message inspection, including NATO- and EU-related targeting rules in the Romanian case. The tooling links support high-confidence Turla attribution, while direct FSB tasking remains an inference at moderate confidence. We also do not have a complete victim count, proof that every matched message was exfiltrated, or a verified account of what operators ultimately received. The defensive implication is still strong: absence of outbound command-and-control traffic cannot clear an Exchange server; transport-process injection, wrapped handlers, anomalous inbound HTTPS, and mailbox exposure must be examined.
Arcus gives us an equally useful distinction between prevention, containment, repair, and recovery. On Arcus’s reported timeline, the bridge rate limit—not the later human intervention—directly blocked roughly $3 million in additional withdrawals. The withdrawal halt contained subsequent exposure, the accounting fix addressed the suspected path, and treasury funding absorbed losses rather than preventing them. Yet the claimed $2.23 million loss, blocked amount, and preservation of user balances have not been independently reconciled transaction by transaction, so those remain company-reported figures.
The common thread is that control boundaries matter more than labels: privileged management reach, trusted mail processing, and withdrawal authorization each became the decisive enforcement point. That takes us naturally to the next comparison—how Outlook and SonicWall represent different identity failures—and then to whether AI-assisted exploitation is genuinely compressing response windows to minutes, or whether that claim outruns the measured evidence.