November 2026 Summaries

Post-quantum cryptography is necessary not only for core Internet protocols such as TLS but also for privacy-preserving abuse-prevention systems like Privacy Pass, whose current RSA- and elliptic-curve-based tokens could be forged by quantum attackers even though their unlinkability remains secure. Cloudflare and collaborators are developing MOLE, a framework for anonymous credentials that can hide both client and issuer identities, enabling equitable moderation and more flexible management of automated traffic. A promising approach uses VOLE-in-the-Head zero-knowledge proofs with multivariate cryptography, particularly UOV signatures, to build post-quantum blind signatures and issuer-hiding endorsements; the PoMFRIT construction demonstrates blind signatures around 4.3 KB, while proposed VOLE-IHAT and VOLE-ACT schemes could support MOLE endorsements and credentials with larger but potentially practical communication costs and millisecond-scale runtimes. Although these designs require further security proofs, cryptanalysis, parameter refinement, and protocol engineering to manage their greater size, the authors argue that post-quantum anonymous credentials for the web may be nearing practical deployment.

AI agents hosted as cloud services increasingly act as intermediaries between users and websites, obscuring the user’s identity, reputation, and accountability while allowing automated activity to operate continuously and at scale. The proposal argues that this shift creates challenges for sites involving fraud disputes, compliance, account sharing, authentication, resource consumption, content monetization, abuse mitigation, and distinguishing legitimate users when traffic from many people is combined through one agent. It advances an “Intermediation Principle” that sites should retain comparable ability to recognize users, apply policies, and assign responsibility whether users act directly or through agents. Human-in-the-loop CAPTCHA checks and broad access permissions are presented as insufficient because they neither establish ongoing user accountability nor resolve shared-agent abuse. Instead, the text calls for open, interoperable standards enabling agent identification, privacy-preserving reputation, graduated user identity disclosure, verifiable claims such as age or cardholder status, and machine-readable site policies, citing efforts including Web Bot Auth, MOLE, OAuth, Privacy Pass, and IETF work. It also announces the Trusted Agent Delegation coalition to coordinate stakeholders and describes Cloudflare tools that let customers identify agent traffic, establish per-agent access policies, and link certain agent activity to hashed user identifiers to mitigate abuse without blocking all users of a shared agent.

Cloudflare has introduced analytics and access controls designed to help site owners identify, evaluate, and manage traffic from personal AI agents acting on users’ behalf. Security Analytics now classifies Personal Assistants as a distinct traffic category and provides details on individual agents, including request volumes, visited paths, status codes, and network origins, while Account Abuse Protection can associate agents with specific user accounts and login or sign-up activity. Site owners can use these insights to distinguish useful customer-directed activity from potential fraud or abuse, including applying targeted protections to an individual account without blocking an agent for all users. Cloudflare has also expanded and renamed its AI Crawler and Agent access policies, allowing overall and agent-specific choices to allow access, block agents on ad-supported pages, or block them entirely. The company says future work will focus on privacy-conscious ways to preserve user context, assess agent reputation, and establish accountability as agents increasingly mediate interactions between users and websites.

Cloudflare has introduced Zero Wire, a public vulnerability intelligence feed within Cloudflare Radar designed to help security teams assess and respond to significant critical vulnerabilities amid faster discovery driven partly by AI. Rather than cataloging every CVE, Zero Wire provides concise briefings on critical Internet-facing software flaws, covering their practical impact, affected deployments, known exploitation, recommended next steps, vendor guidance, supporting evidence, and applicable Cloudflare mitigations. Briefings are organized into “Needs attention” for issues requiring current assessment and “Reference” for stabilized guidance, while search, technology tags, update histories, and last-checked dates help users track developments. The service also offers AI-agent prompts for read-only infrastructure assessments and distinguishes evidence of attempted exploitation, exposure, and compromise. Cloudflare emphasizes that WAF protections can serve as compensating controls but do not replace vendor patches or other remediation, and that the resource is available to both Cloudflare customers and noncustomers.

Cloudflare redesigned its DNS Zone Builder pipeline to handle growing volumes of DNS and zone-setting changes without repeatedly rebuilding entire zones that can contain tens of millions of records. The new system uses a targeted build planner to identify the smallest safe unit of work, such as specific DNS names, apex records, shared metadata, or a purge, while retaining full builds for uncertain or inherently zone-wide changes and applying special safeguards for very large zones. A unified scheduler merges pending work for the same zone, preserves Kafka delivery guarantees, prevents busy zones from monopolizing workers, and verifies current database state before writing updates to Quicksilver, Cloudflare’s distributed key-value store for authoritative DNS data. Because zones with the same domain name can share metadata, the system rebuilds that data from fresh publication state and coordinates rollout by domain name to avoid conflicting writes. Cloudflare validated the redesign through byte-level equivalence tests, production mirror comparisons, and staged deployment, finding that 92.95% of builds avoided complete-zone rebuilds, peak memory use fell by 85%, Kafka consumption became steadier, worst-case p99 completion-delay spikes dropped from 14.6 minutes to 1.21 minutes, and subsequent reductions in scheduling and replication buffers cut median completion delay from about 7.7 seconds to 3.3 seconds.