No PriorsDemocratizing Security for the Agentic Web with CloudFlare CTO Dane Knecht
CHAPTERS
- 0:00 – 0:58
AI-driven “vuln apocalypse”: why security became the #1 priority
Dane frames AI as both the accelerant behind a rapid increase in discovered vulnerabilities and the tool that can help fix them. He explains how the industry shifted from “get everyone using AI” to urgently defining what AI security means across data leakage and intrusion prevention.
- •AI is simultaneously creating and solving a security crisis
- •CTO priorities anchored on Cloudflare’s mission: a better, safer internet
- •AI adoption initially outpaced security planning (security as afterthought)
- •AI security spans both data exfiltration risks and classic break-in prevention
- 0:58 – 3:30
From annual Log4j moments to weekly critical bugs
The conversation zooms in on the changing tempo of internet security incidents. Dane argues that vulnerabilities comparable to past “all-hands” events now appear with startling frequency, driving a near-term reallocation of engineering time toward security.
- •Critical vulnerability cadence has accelerated dramatically
- •Cloudflare treats protecting free users as a public-good obligation
- •Prediction: a majority of engineering time goes to security in the next ~18 months
- •AI will eventually “survey” much of the codebase ecosystem and reduce unknowns
- 3:30 – 4:59
Brittle foundations: randomness, keys, and AI-assisted cryptanalysis
Dane uses recent examples to illustrate how AI plus modern tooling can uncover deep security flaws, including weak cryptographic practices. He emphasizes that models don’t sleep, and the combination of better models and better “harnesses” is raising attacker capability fast.
- •Example of catastrophic weakness: cryptographic material seeded insecurely
- •Cloudflare’s focus on strong randomness (e.g., lava lamps) as a principle
- •AI + automation scales vulnerability research beyond human red teams
- •Improvement isn’t only models; it’s how humans operationalize them (harnesses)
- 4:59 – 6:12
Who’s attacking now: from nation-states to “agent kiddies”
Sarah asks who is behind the surge in exploitation. Dane notes that some high-impact findings are coming from non–nation-state actors, including younger hackers, incentivized by bug bounties and empowered by agentic tooling.
- •Some major vulns are discovered by individuals, not nation-states
- •Agents turn “capture the flag” curiosity into scalable real-world discovery
- •Bug bounties and financial incentives amplify participation
- •Advanced capabilities are being democratized to the masses
- 6:12 – 7:28
Doomsday vs. adaptation: zero trust, industry coordination, and resilience
Sarah raises fears about needing to air-gap critical systems and widespread infrastructure failure. Dane is less doomsday, arguing that crisis drives rapid collaboration across competitors and that the shift from castle-and-moat to zero trust improves baseline readiness.
- •In crises, competitors collaborate to protect critical infrastructure
- •Zero trust improves posture versus older perimeter-based security
- •Tightening device/flow controls is more realistic than fully air-gapping
- •Humans and organizations adapt quickly when pressure is high
- 7:28 – 9:46
Why security spend will spike—and why it won’t last forever
Dane predicts an emergency-driven surge in both offensive (red team) and defensive investments. He expects the spike to be a transitional period, after which productivity rebounds as tooling and processes mature.
- •Security budget shift is driven by emergency, not gradual planning
- •Investment rises on both red team and defensive preparedness
- •Expectation: a distinct ~18-month “security-heavy” phase
- •Longer-term: productivity increases once the backlog stabilizes
- 9:46 – 11:43
Building more secure software: reduce attack surface and fight AI verbosity
Dane shares a pragmatic philosophy: the fastest security win is simplification—shrinking code footprint and surface area. He notes today’s models often generate unnecessary complexity, making human judgment and disciplined engineering critical.
- •Primary principle: reduce attack surface by simplifying systems
- •Formal verification, safer languages, and secure runtimes help—but later
- •Current-gen AI tends to be verbose/complex, which can harm security
- •Humans must enforce simplicity and secure design standards
- 11:43 – 12:49
Cloudflare’s secure-build playbook: Codex, principles, and adversarial reviews
Dane explains how Cloudflare codifies secure engineering via internal documentation (“Codex”) distilled into enforceable principles. Agents can follow these rules throughout review pipelines, but adversarial reviews and security teams remain essential.
- •Security must be addressed at all levels: engineers, security teams, and process
- •Cloudflare’s “Codex” aggregates secure/available software best practices
- •Principles get distilled into “Ten Commandments” agents can apply
- •Maintain adversarial reviews (performance/security runs) alongside agent output
- 12:49 – 14:39
The messy middle: humans and agents sharing the same web rails
They discuss how agent-driven traffic is reshaping the internet. Dane argues agents are inefficiently parsing the human web and that the current state forces costly translation layers, motivating a split between human and machine experiences.
- •Agents consuming the human web is inefficient and token-expensive
- •Human web includes ads and pacing; agents operate at massively higher speed
- •Likely emergence of parallel experiences: human web vs machine web
- •Need to “meet agents where they are” rather than forcing one interface
- 14:39 – 17:23
New economics for content: microtransactions, CPM equivalents, and payment rails
Dane proposes an agent-compatible monetization model where agents can pay for content at rates analogous to ad CPMs. This requires internet-scale micropayment infrastructure and could meaningfully expand creator opportunities.
- •Agents should compensate publishers rather than free-riding via crawling
- •Model: agent decides to read, then pays a CPM-like rate
- •Requires new payment rails supporting millions of microtransactions/sec
- •Agents may finally make long-promised micropayments practical
- 17:23 – 19:09
Cloudflare roadmap: monetization gateway + “Signals” to reduce crawling costs
Dane describes Cloudflare products aimed at enabling these new models: a monetization gateway with agent wallets and stablecoin support, and a collaboration with OpenAI called Signals to help agents discover new content without constant crawling.
- •Monetization gateway: simple paywalls/microtransactions for sites
- •Providing wallets for agents and stablecoin-enabled settlement
- •Signals (with OpenAI): proactive discovery of new content vs repeated crawls
- •Benefits both AI labs (efficiency) and site owners (lower infra burden)
- 19:09 – 20:03
Defining success for creators: Cloudflare pays publishers, expands creation
Sarah asks what would make the model become “default.” Dane says success is when monetization covers creators’ infrastructure costs and meaningfully supports livelihoods—potentially increasing the number of viable content creators.
- •Success metric: creators get paid (Cloudflare paying them, not vice versa)
- •Revenue should cover infrastructure and support livelihoods
- •Goal: increase the number of creators beyond the last internet era’s growth
- •Shift from ad-only economics to agent-native compensation
- 20:03 – 24:01
Cloudflare OS: a secure internal AI platform that becomes a reference architecture
Dane explains Cloudflare OS as an internal system for employees to use AI safely with governance over data flows. It also encodes “how Cloudflare works” for agents, and is released externally as a forkable reference architecture built on Cloudflare’s developer platform.
- •Built to prevent internal/customer data leakage while enabling AI usage
- •Encodes company practices so agents can operate with Cloudflare-specific context
- •Released as reference architecture; customers can fork a “company OS”
- •Demonstrates how to build end-to-end on Cloudflare’s platform primitives
- 24:01 – 35:11
Becoming AI-first: restructuring, workflow reinvention, token governance, and hiring for curiosity
They cover Cloudflare’s operational changes: restructuring to redesign workflows around AI, emphasizing bottom-up ownership, and evolving token budgeting through smart routing and lightweight guardrails. Dane also highlights an intern-heavy strategy and a hiring filter centered on curiosity and real AI use.
- •Restructuring aimed to rebuild processes, not just speed up existing teams
- •Example workflow: PM writes cohort + email; AI rewrites in voice; manager approves; auto-send
- •Token budgeting should be intelligently routed; humans shouldn’t pick models per task
- •Governance via usage notifications/escalations; optimize for output shipped, not token vanity
- •1,000-intern strategy to inject AI-native behavior; curiosity as a core hiring signal
- 35:11 – 41:23
Closing loop on security: scale red teams, balance shipping, open vs closed models, and democratized defense
In the final section, Dane explains how Cloudflare will tackle legacy security debt by scaling adversarial discovery (more red teams) while continuing to ship. He argues open access to capable models is inevitable and necessary, and positions Cloudflare as a way to democratize security for organizations that can’t afford large security teams.
- •Plan: expand from one red team to multiple teams focused on discovery
- •Processes matter, but crises force automation and scalable patching
- •Security must coexist with shipping—no feature freeze; AI can speed both
- •Open vs closed: capabilities will spread anyway; restricting access harms defenders
- •For smaller orgs: put sites/infrastructure behind Cloudflare to inherit protections
- •Cloudflare frames security/performance as utility-like democratization (including free tier)
