The clearest line across these cases is that exposure, targeting, and compromise are not interchangeable. Alex has separated the 216 PHP organizations identified in reconnaissance from the single documented government-server case involving code execution, web-shell persistence, NTLM hashes, MFA secrets, relay activity, and directory-data theft. The Lazarus attribution remains moderate-confidence, and that full kill chain cannot be projected onto the other potential targets. Moving an organization from “exposed” to “suspected compromise” requires exploit-shaped activity plus a host or network consequence; confirming web-shell use or credential theft requires the corresponding artifact, execution evidence, or credential-access telemetry.
The same evidentiary discipline changes the response elsewhere. For BeyondTrust CVE-2026-102764, pre-patch exploitation means an affected Remote Support appliance can no longer be assumed to be a trustworthy broker. It does not, by itself, establish that every technician, endpoint, federation token, or downstream credential was compromised. Isolation, preservation, remediation, retrospective hunting, and session review are justified broadly; revocation and credential rotation should then follow the actual exposure path and suspicious activity rather than an unsupported assumption of universal compromise. In the Plex2Shell/QNAP case, there is strong support for an exploitable four-flaw chain, public proof of concept, and active exploitation, but only moderate support for confirmed QNAP compromises and no incident-level proof that every attacker used the complete published chain. Opportunistic mass exploitation is more defensible than a coordinated campaign attribution.
Sara has also kept the refrigeration incidents from collapsing into one narrative. The roughly 700–750-site French cluster carries a reported Shahid Kaveh attribution, while the later nine-country wave does not. The likely safety concern is supervisory access with authority to alter defrost schedules, setpoints, or alarms, creating spoilage, relocation, equipment-damage, and possible fire risks; the initial access mechanism and some claimed physical consequences remain unverified.
We now carry that same framework into the next set: distinguish mechanism from impact, invalidate only the trust boundaries the evidence supports, and treat attribution, scope, and downstream harm as separate claims. That will be especially important across the Kava Lend loss, the athlete-data breach clock, Cursor’s developer supply-chain compromise, and claims that AI is creating a genuinely new exploit class.