Send us Fan Mail
CISSP Domain 2.6 — data security controls, compliance requirements, and the three states of data, built on an AI hosting exposure that encryption could never have reached, because the data was exposed while it was being used.
Start free: https://www.cisspcybertraining.com/free-cissp-self-study (360 practice questions, a Domain 1 diagnostic, weekly questions by email, and the slides from this episode)
One hundred and fifty-five answers went to the wrong people. Nobody broke in.
In late August, Cloudflare launched the GLM-5.3 model on Workers AI. The model advertised a context window of just over one million tokens. The memory pool behind it — the prefill KV cache — held about six hundred and eighty thousand. The pool was smaller than the thing it was advertised to hold.
Picture a shared workbench with a scratch pad on it. You sit down, you do your working-out, you get up and leave. In a well-run shop somebody wipes that pad before the next person sits down. That pad is memory, and wiping it is the control.
When a request came in too big for the pad, the system marked it rejected, which was correct, and then passed the pad downstream anyway. The next stage read it, found the previous request's working still on it, and handed that back as a successful answer. Not an error. An HTTP 200.
155 responses were affected, every one from a prompt over the pool's limit. Cloudflare said only that a small number of customers were involved and published no customer count. It found no evidence of deliberate exploitation. This was an exposure — nothing was taken, nobody broke in — and the misconfigured value and the serving-engine behavior both sat in the hosting stack, not with the model's maker. The advertised context length itself was accurate.
Security researcher Gadi Evron of Knostic reported it through HackerOne. Cloudflare began verifying it the next morning and reproduced it from two independent accounts within twenty-five minutes of declaring the incident — and the model was offline about fifteen hours after the report landed. Root cause thirty-five minutes later, corrected configuration seventeen minutes after that. It raised the pool to roughly twice the advertised window, moved retirement of rejected requests to the admission queue before any cache transfer, added a deployment check comparing pool size to context length, and published the whole timeline itself.
Now the part that turns it into exam content. Encryption at rest and encryption in transit were never going to touch this one. Not because anything was wrong with them. Because neither one has a single thing to say about data while it is being used.
And they knew. Cloudflare knew before launch that the two numbers disagreed. It assumed the engine would handle it cleanly, and what it validated was the status code and the performance — not the answer coming back. The engine said 200 OK, so it looked fine. It was not a secret. It was an assumption nobody had validated, standing in for a control nobody owned.
A control is not a control because of what it is called. It is a control because of what fails when it is wrong.
In this episode:
• The three states of data — at rest, in transit, in use — and why the third is the one you get tested on. At rest and in transit are cryptography questions. In use is an isolation question, and cryptography alone does not reach it. Confidential computing is the exception.
• The data-in-use control vocabulary: memory isolation, zeroization between requests, object reuse controls, confidential computing, homomorphic encryption and admission control. Object reuse is the classic ISC2 phrasing for this failure shape.
• Scoping and tailoring, and the precision trap. Scoping is not a step before tailoring — it is one of the six tailoring actions NIST SP 800-53B names, and an option that sequences them is the distractor.
• Standards selection and baselines: ISO 27001, NIST SP 800-53 and 53B, CSF 2.0, CIS, the Cloud Controls Matrix, PCI DSS and SOC 2, taught by differentiator rather than contents. A baseline is a floor, not a fit.
• Data protection methods — DRM, DLP, CASB, tokenization and masking — and why not one of them would have caught this. Every one reads a content flow.
• Multi-tenancy and shared responsibility, and the control each side assumed the other already owned.
• The crossover, named out loud: "should we accept this?" is Domain 1. "How is it enforced?" is Domain 3. "What does this data require?" is 2.6, and that is this episode.
• Four CISSP practice questions and the trap inside each one. Question two is the one that gets people.
The line to carry into the exam room: Domain 2.6 does not ask you to list encryption types. It asks which controls this data, in this state, in this environment actually requires — and then whether anybody owns them as requirements.
LINKS
Free CISSP Essentials:
https://www.cisspcybertraining.com/free-cissp-self-study
All episodes and the written walkthrough:
https://www.cisspcybertraining.com/podcast
The CISSP Sprint Cohort — 8 weeks, live, paced to your exam date, with an accountability partner:
https://www.cisspcybertraining.com/cohort-sprint-page
SOURCES
Primary — Cloudflare's own incident writeup, published late October 2026:
"Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang"
https://blog.cloudflare.com/kv-cache-exposure-remediation
Reported by security researcher Gadi Evron, Knostic, via HackerOne.
Shon Gerber, CISSP — former Commander, 177th Information Aggressor Squadron. Former CISO, fractional CISO, founder of CISSP Cyber Training.
Gain exclusive access to 360 FREE CISSP Practice Questions at FreeCISSPQuestions.com and have them delivered directly to your inbox! Don’t miss this valuable opportunity to strengthen your CISSP exam preparation and boost your chances of certification success.
Join now and start your journey toward CISSP mastery today!
