The month that broke the CVE record
In July 2026 Microsoft shipped the largest Patch Tuesday in the program's history: 570+ CVEs at once (622 by some counts that include the advisories), three of them zero-days already under exploitation. It is not an isolated outlier month. It is the peak of a trend the industry itself started to name: AI is generating (and helping find) flaws faster than teams can patch them.
Among the month's criticals, a portrait of the new risk map: an elevation of privilege in Azure OpenAI (CVE-2026-45499, CVSS 9.9), secret reading in IBM Langflow (CVE-2026-10134, an AI orchestrator), and command injection in Arista VeloCloud (CVE-2026-16812, CVSS 10.0) already exploited in the wild. Each new AI layer in the stack became a new attack surface.
Why the volume exploded
Two forces at once. First, AI tools accelerate software creation, so there is more code, faster, with more embedded flaws (2026 measurements put 60%+ of AI-generated code shipping with some vulnerability). Second, AI also accelerates discovery: fuzzers and analysis agents find bugs at a scale that used to depend on a human researcher. More fragile code plus more finding capacity equals a flood of CVEs.
For the defender, that changes the patching math. When dozens of CVEs shipped per month, you could try to fix them all. With hundreds per month, "patch everything" stopped being a strategy and became a way to drown.
The founder mistake: treating CVEs as a to-do list
The reflex is to look at the list, see the scary number, and either freeze or try to apply everything. Both fail. Most of a month's CVEs do not affect you: you do not use that component, you are not exposed, or the flaw is not reachable in your context. Spending equal energy on all of them dilutes the effort away from what matters.
What matters is the intersection of three things:
- Reachable: can an attacker get to that surface from the internet or an ordinary account?
- On your revenue or data path: does the flaw touch login, billing, customer data, or what sustains the sale?
- Exploited in the wild: is there already an exploit running, like the month's three zero-days?
A CVE that hits all three is a fire and jumps the queue. One that hits none can wait for the normal update cycle.
How to prioritize without becoming a hostage to the number
- Keep an inventory of what you actually use, direct and transitive dependencies. You cannot prioritize what you do not know you have.
- Alert by severity AND exposure, not just CVSS. A 9.8 in a component you do not expose is less urgent than a 7.5 on your payment route.
- Treat an exploited zero-day as an interruption: it is not in the queue, it is the queue.
- Do not confuse "green scanner" with secure. The scanner sees a known CVE in a dependency, not the logic flaw AI left in your code (IDOR, plan bypass, cross-tenant leakage).
Where RET fits
The CVE flood is noise; the real risk is what is reachable and exploitable in your product. That is what RET measures: Risk Review reads the flows that touch access, money and data and separates what is a fire now from what is a list; the Manual Pentest proves what an attacker actually reaches. In a 570-CVE month, knowing which five affect you is worth more than the whole list.
In a world that generates hundreds of CVEs a month, the edge is not patching faster, it is knowing which few truly affect you.
