Patch these two now: CoSnitch in Copilot, and CVE-2026-65346 in Apple ImageIO
One sat known for eight months. The other needs no interaction beyond receiving an image. Neither is theoretical.
Two disclosures landed this week that belong on a change board today rather than in a backlog. They are different in kind, and the second is worse than it sounds.
CVE-2026-65346 — integer overflow in Apple ImageIO
An integer-overflow flaw in the ImageIO framework affects iOS, iPadOS and macOS, and can lead to arbitrary code execution when a malicious image is processed. It was reported by Meta’s Red Team.
ImageIO is the shared decoding path. It is what renders a thumbnail in Messages, a preview in Mail, an avatar in a chat client, an attachment in a calendar invite. Which means the interaction required to reach the vulnerable code is frequently none — a message arrives, the OS decodes a preview, the parser runs on attacker-controlled bytes.
That class of bug — memory corruption in a system image parser reachable without a tap — is the historical backbone of commercial spyware delivery chains. Treat it as an active exploitation risk, not a hypothetical.
- Patch mobile fleets first; the message-preview path is the exposure.
- Do not rely on user training. There is no user action to train against.
- If you manage devices for people at elevated risk, escalate this above normal patch windows and consider Lockdown Mode where it is tolerable.
CoSnitch — data exfiltration via Copilot
Microsoft patched a critical vulnerability in Copilot, tracked publicly as CoSnitch, that allowed data exfiltration. Reporting indicates it was known for roughly eight months before the fix shipped.
The eight months is the story. A data-exfiltration path in an assistant with tenant-wide document access is about as sensitive as enterprise software gets, and the exposure window was not a disclosure-coordination lag — it was three quarters.
The deeper issue is that assistants integrated across a tenant collapse the blast radius of a single bug. A classic exfiltration bug leaks what the compromised user could reach. A bug in a system with retrieval across the whole tenant leaks what the assistant could reach, which is a much larger set by design. That is not an argument against deploying them. It is an argument for treating them as data-plane infrastructure with corresponding patch urgency and audit logging, rather than as a productivity feature.
The week’s breaches, for calibration
Three data points on where losses are actually coming from:
| Target | Vector | Scope |
|---|---|---|
| France, Education Ministry (Siecle) | Claimed by a group calling itself “ZeroBytes” | Millions of student and teacher records: contact details, addresses, school records |
| France, tax authority | Not disclosed | 678,000 individuals |
| Quest Apartment Hotels | Third-party service provider | Names, emails, addresses, dates of birth, historical records pre-June 2025 |
Note the pattern in the third row, because it is the most common one and the least defended: the vulnerability was not in the company that lost the data. Third-party and supply-chain exposure is where a mature internal patching programme still loses, and the Quest records dated from before June 2025 — data retained past any plausible operational need, which converted a vendor breach into a much larger disclosure.
Retention policy is a security control. Data you deleted on schedule cannot be exfiltrated by your vendor’s bug.
The short version
- Today: Apple ImageIO updates across iOS, iPadOS and macOS. Zero-interaction RCE in a system parser.
- Today: Copilot updates. Assume the exfiltration path was reachable for the full window and review assistant audit logs accordingly.
- This quarter: inventory third-party processors and enforce retention deletion. Two of three breaches above were made worse by data nobody needed to be holding.