Building Enterprise Frontier Safeguards With Our Customers  🔗 **Source:** [Building Enterprise Frontier Safeguards With Our Customers](https://www.youtube.com/watch?v=FoteuzPpx7E) > [!note]+ Document Notes > tags: [anthropic, claude, enterprise-ai, ai-safety, privacy-architecture, security-logging, visa, uber, customer-testimonial] > author: Claude (Anthropic YouTube channel) > source: YouTube — @claude > reference: https://www.youtube.com/watch?v=FoteuzPpx7E > date created: 2026-09-01 21:35 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > > [!warning] Thin source — ~220 words of testimonial, speakers unidentified > > The transcript is a stitched customer-testimonial reel of roughly two minutes with no speaker > > names, titles, or on-screen text captured. Two companies are identifiable from the speech > > itself (Uber, Visa); Anthropic is named once. Channel confirmed as Anthropic's "Claude" > > account via YouTube oEmbed. "Enterprise Frontier Safeguards" is the product name in the title; > > the transcript never defines it. Where this note describes what the product *does*, it is > > restating the speakers' claims, not verifying them. > > # 🏷️ Title: Building Enterprise Frontier Safeguards with our customers > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - Source: https://www.youtube.com/watch?v=FoteuzPpx7E (Anthropic "Claude" channel, published 2026-09-01, ~2 min) - No links in the stored document (verified: zero `<a>` tags). ## 📝 Summary **What this is.** A customer-testimonial reel for an Anthropic enterprise offering called "Enterprise Frontier Safeguards." Unnamed speakers from Uber and Visa explain why they trust it. The clip's one technical passage — from the Visa speaker — describes an architecture rather than a promise, and that passage is the only part of the reel that carries information. **The architecture claim, as stated.** Security and privacy controls live "at the architectural level, not just policy level." Concretely: (1) the customer's data stays in the customer's cloud under the customer's control; (2) the customer has continuous access to the signals; (3) logs never leave unless the customer chooses; (4) safety review is machine-only — no human at Anthropic reads customer content; (5) what the review emits is "intentionally limited to defined findings, not customer content." **Why that passage matters.** Points 4 and 5 describe a specific privacy pattern: a monitoring system whose output is a fixed schema of findings rather than the material it inspected. That's the same shape as a linter that reports rule IDs instead of source lines, or a DLP scanner that flags "credit card detected" without exfiltrating the number. If it works as described, it lets a bank run frontier-model review over regulated data without the reviewer becoming a data recipient. **Why the rest doesn't.** The Uber line ("bone-deep focus on safety") and the Visa framing (billions of consumers, 160M merchants, 15,000 banks, 130 years of trust) are brand credentials. They establish that these customers care; they don't establish that the product delivers. "Visa cannot do it ourselves" is the reel's sales thesis — frontier labs supply "ingredients" incumbents can't build — and it's asserted, not shown. **Stakeholders.** Anthropic (positioning against other labs on enterprise trust); regulated-industry buyers (finance, mobility); privacy and security teams who will need to verify the architecture; anyone whose payment or ride data flows through these systems. **Risks.** The clip's own framing — "the stakes are higher" now — is correct and unelaborated. The architectural claims are unfalsifiable from a testimonial; verifying them needs a technical writeup, a SOC 2 / ISO 27001 scope statement, or a data-flow diagram, none of which the clip references. **Bottom line.** Discard the testimonial, keep the design pattern. "Machine-only review, output limited to defined findings, logs under customer control" is a three-clause architecture worth naming and checking against whatever documentation exists. ## 📖 Main Content The reel runs in three movements. **Credentialing (Uber, then Visa).** Uber's speaker asserts a shared safety culture with Anthropic. Visa's speaker recites scale — billions of consumers, 160 million merchants, 15,000 banks — and duration — 130 years — then frames trust as the asset at risk: "That trust took decades to earn, and I don't intend to spend it carelessly now that the stakes are higher." **The architecture (Visa).** The only substantive segment. Controls are built in, not bolted on ("architectural level, not just policy level"). Data residency: "store in our cloud with our control." Observability: "continuous access to the signals." Egress: "The logs are under our control. They don't go anywhere else unless we want them to." Review: "machine-only." Output: "intentionally limited to defined findings, not customer content." **The thesis (Visa).** Incumbents can't build this alone; frontier labs supply the "ingredients." The safeguards unlock use of AI "in parts of the business that we wouldn't have been able to do so before" while remaining "responsible stewards of the capabilities of these models." Nothing in the reel says what a "finding" is, who defines the schema, whether findings can be appealed, what the review is looking for, or where the model itself runs. ## 🗒️ Notes - [00:00] Uber: "bone-deep focus on safety" — shared-values framing, no content. - [00:06] Visa: scale credentials — billions of consumers, 160M merchants, 15,000 banks. - [00:14] "Trust is paramount to our brand. It's literally our business model for 130 years." - [00:20] "That trust took decades to earn, and I don't intend to spend it carelessly now that the stakes are higher." — the reel's best line; also its only acknowledgment of risk. - [00:28] "new security and privacy capabilities at the architectural level, not just policy level" — the enforcement-ladder claim. - [00:35] Data residency and control: "store in our cloud with our control… continuous access to the signals." - [00:42] Egress: "The logs are under our control. They don't go anywhere else unless we want them to." - [00:48] "The review is machine-only. What comes out is intentionally limited to defined findings, not customer content." — the privacy pattern. - [00:56] "Visa cannot do it ourselves. We need the frontier labs like Anthropic helping us to provide the ingredients" — the sales thesis. - [01:08] Safeguards as enabler: AI "in parts of the business that we wouldn't have been able to do so before." - [01:18] Close: "responsible stewards of the capabilities of these models." ## 📚 References **Stated by the source:** none. No documents, standards, or URLs are named. **What would substantiate the claims (not provided):** - An Anthropic technical post or docs page defining "Enterprise Frontier Safeguards" and the findings schema. - Data-flow documentation showing where inference runs relative to the customer's cloud (the clip says logs stay; it doesn't say the model does). - Third-party attestation scope (SOC 2 Type II, ISO 27001) covering the machine-only review path. - Visa's and Uber's own disclosures, if any, on AI-vendor data handling. **Named entities:** Anthropic, Uber, Visa. Speakers unnamed in the transcript. **Link audit.** Zero `<a>` tags in `html_content`. Nothing to check. ## 🗣️ Paraphrase An Uber executive says Anthropic shares Uber's safety obsession. A Visa executive says Visa protects billions of people and has spent 130 years earning trust it doesn't intend to waste now that AI raises the stakes. Working with Anthropic, Visa got controls built into the system rather than written into a policy: Visa's data stays in Visa's cloud, Visa can see all the signals continuously, logs go nowhere without Visa's say-so, the safety review is done by machines rather than people, and what the review reports is a fixed set of findings rather than any actual customer data. Visa couldn't have built this alone; it needs frontier labs to supply the pieces. The result is that Visa can now use AI in parts of the business it previously couldn't touch, while still being careful with what these models can do. ## 📰 Newsletter **The two sentences worth keeping from Anthropic's enterprise reel** Anthropic put out a customer-testimonial clip this week for something called Enterprise Frontier Safeguards. Most of it is what you'd expect — Uber says safety, Visa says 130 years of trust — but buried in the Visa segment are two sentences that describe an actual design: "The review is machine-only. What comes out is intentionally limited to defined findings, not customer content." That's a pattern, and it's one WordPress developers already use without naming it. Your linter reports rule IDs, not your source. Your uptime monitor reports "down," not your database. A good security scanner tells you "hardcoded secret on line 40" and doesn't ship the secret anywhere. The Visa version applies it to frontier-model review over payment data: the model looks, a schema of findings comes out, the content stays home. The other line worth stealing: "architectural level, not just policy level." A policy is a rule someone can ignore. An architecture is a rule that isn't offered as a choice. If you've ever written a CONTRIBUTING.md that nobody reads and then added a pre-commit hook that everybody obeys, you know the difference. Everything else in the clip is credentials. Brand scale isn't evidence a product works; it's evidence a customer had a lot to lose and signed anyway. Wait for the docs. - Watch: https://www.youtube.com/watch?v=FoteuzPpx7E ## ✅ Action Items **DevRel** - "Architecture, not policy" is a one-line frame for the enforcement-ladder argument Courtney already makes in her own tooling docs (write gates over instructions). Cite this clip as an external instance when the point needs an industry voice. `#action/h3` **OSPO** - Not applicable to this source. No open-source licensing or governance content. **Community** - Not applicable to this source. **WordPress teams** (Core, Core Performance, Core AI, Core Program Team, Mobile, Playground, Community, Training, Marketing, Support, Design, Documentation, Photos, TV, Polyglots, Accessibility, Meta, CLI, Hosting, Tide, Openverse, Themes, Plugins) - **Core AI:** if the AI team ever specifies a review or moderation feature, "output limited to defined findings, not content" is the privacy contract to write down first. `#action/h2` - **Plugins / Tide:** the same pattern describes what a plugin-review scanner should emit: findings schema, never plugin source to a third party. Check whether the current Plugin Check tooling documents its egress. `#action/h3` - **Hosting:** hosts positioning "AI security review" offerings should be asked the same three questions — where do logs live, is review machine-only, what does the output contain. `#action/h3` - All other teams: not applicable to this source. **The WPOCC** - Not applicable to this source. **FAIR** - Not applicable to this source. ## 🧪 Hypotheses 1. **"Machine-only review with findings-only output" is materially different from standard vendor DPAs.** Most AI-vendor agreements permit human review of flagged content. If Anthropic's enterprise tier contractually excludes it, that's checkable in the published terms. Falsifiable by reading the DPA. 2. **The model does not run in the customer's cloud.** The clip says *logs* stay; it conspicuously doesn't say *inference* does. Prediction: the architecture is customer-hosted logging around vendor-hosted inference. Falsifiable by any data-flow diagram. 3. **Enterprise buyers weight "architectural vs policy" language more than attestation logos.** A survey of regulated-industry buyers would show the phrase tests better than "SOC 2 certified." Testable. 4. **Testimonial reels from frontier labs converge on the same three credentials** — scale, duration, shared values — regardless of product. Checkable against the other lab channels' enterprise clips over the next quarter. 5. **The findings schema is narrower than customers expect.** If "defined findings" means a fixed enumeration of safety categories, customers will discover use cases the schema doesn't cover within the first year. Testable only with access to the schema. ### 📊 Methodology **Genre: customer testimonial reel.** Edited excerpts from at least two interviews, sequenced credentials → architecture → thesis. No method in the research sense; the evidentiary standard is "a trusted brand said so." **What is methodologically strong.** The architecture passage is unusually specific for the genre. Five concrete properties (residency, observability, egress control, machine-only review, findings-only output) is more than most enterprise marketing offers, and each is in principle verifiable. **What is weak.** Every claim comes from a customer with an announced commercial relationship. No independent party. No definitions — "finding," "signal," "safeguard" are all undefined. Speakers are unnamed in the captured transcript, so even the attribution to Visa and Uber rests on self-identification in the speech ("At Visa, we…"). The reel offers no failure case, no limit, no "what it doesn't do." **How this analysis proceeded.** Separated the one information-bearing passage from the credentialing around it, restated its five claims as checkable properties, and noted the one omission (where inference runs) that the wording seems designed to skate past. ### ❓ Questions - **Where does inference run?** Logs in the customer cloud is stated; model in the customer cloud is not. This is the question every security team will ask first. - **What is a "defined finding"?** A category label, a severity, a pointer? Who controls the schema — Anthropic, the customer, or both? - **Is "machine-only" contractual or operational?** A policy can be reversed; a contract clause has to be renegotiated. - **Who are the speakers?** Names and titles would let a reader check whether they own the systems they're describing or the relationship they're promoting. - **What does "the stakes are higher" refer to?** Model capability, regulatory exposure, attack surface? The line does real work and is never unpacked. - **Why Uber and Visa?** Mobility and payments are both high-volume, PII-dense, and regulated. Is that the product's target profile, or just who agreed to be filmed? ### 🔮 Future Research - **Locate the product documentation.** Search Anthropic's site and docs for "Enterprise Frontier Safeguards" and reconcile the five spoken claims against written ones. Any gap is a finding. - **Compare vendor DPAs on human review.** Anthropic, OpenAI, Google, Microsoft enterprise terms — which permit human review of flagged content, under what conditions. A table. - **Catalog "findings-only output" precedents.** DLP scanners, SAST tools, email-security gateways, and content-moderation APIs all emit schemas rather than content. What did each learn about schema drift and appeal paths? - **Architecture-vs-policy in Courtney's own enforcement ladder.** Her deep-read tooling doc ranks write gates and hooks above written instructions with the line "a rule enforced on one write path is not enforced." The Visa passage is an industry-scale instance of the same ranking. Worth a cross-link, not a new note. - **Testimonial-reel structure across labs.** Track the next quarter of enterprise clips from the major labs; test Hypothesis 4. ### 💡 Implications **Practice.** For anyone specifying an AI review or moderation system: write the output schema before the prompt. "What may leave the boundary" is the privacy contract; everything else is implementation. **Policy.** Regulators evaluating AI-vendor arrangements in finance could adopt the five properties as a checklist — residency, observability, egress control, machine-only review, findings-only output — and ask vendors to attest to each. **Security.** Findings-only output shrinks the blast radius of a vendor breach: the attacker gets categories, not content. It also creates a new target — the schema itself, and the model's judgment about what fits it. **Education.** A compact teaching example for "policy vs architecture" — the two-sentence Visa passage carries the whole distinction. **Funding.** Not applicable to this source. No pricing or commercial terms are discussed. ## 🧠 Learning Styles Assessment **Estimated reading level of the transcript:** Grade 10–11 (corporate register, abstract nouns — "architectural level," "responsible stewards," "capabilities"). | Learner type | Fit | Notes | |---|---|---| | **Visual** | Weak | Talking heads, presumably with lower-thirds the transcript doesn't capture. No diagram of the architecture it describes — the single most useful visual the clip could have had. | | **Aural** | Moderate | Clear speech, two distinct speakers, but the content is abstract and fast. | | **Verbal** | Moderate | Corporate diction; the one dense passage rewards re-reading, the rest doesn't. | | **Kinesthetic** | Weak | Nothing to do. | | **Logical** | Moderate | The architecture passage is a five-property list a logical learner can extract; the credentialing around it is noise. | | **Social** | Weak | No interaction between speakers; sequential monologue. | | **Solitary** | Moderate | Fine to consume alone; there's little to discuss. | | **Naturalistic** | Weak | Not applicable to the content. | ## 📊 Accessibility Assessment #### Alternative text and non-text content **Not applicable to the source's own assets, with one missed opportunity.** Talking-head footage carries no informational imagery. But the clip describes an architecture in words alone; a diagram would have helped every viewer and, with proper alt text, would have satisfied **SC 1.1.1** trivially. Its absence isn't a WCAG failure — there's nothing to describe — but it's a comprehension cost. #### Keyboard navigation **Not applicable; inherited from YouTube.** **SC 2.1.1** is the platform's. #### Contrast and readability **Not assessable from the transcript.** Lower-third name/title graphics, if present, would need frame inspection against **SC 1.4.3**. Flagged unassessed — and relevant, since the speakers' identities exist only there. #### Text resizing and reflow **Not applicable to captured content.** Any name cards are baked into frames and can't reflow (**SC 1.4.4 / 1.4.10**), which matters here more than usual because the transcript has no speaker attribution at all. #### Media and captions **Partially met.** Captions exist and the speech is clean. But speaker changes are unmarked in the transcript and the speakers are never named in speech — a deaf viewer relying on captions receives an unattributed monologue that switches company mid-stream. **SC 1.2.2** is met at the word level and fails at the attribution level; speaker-labeled captions would fix it. **SC 1.2.3 Audio Description:** not required, no visual information is conveyed that the audio omits — except the speakers' names, which is the recurring defect. #### Forms and interactions **Not applicable.** No controls. #### Standards reference WCAG 2.2 Level AA. **Met:** 1.1.1 (nothing to describe). **Partially met:** 1.2.2 (accurate words, no speaker labels). **Not applicable:** 1.2.3, 1.4.4, 1.4.10, 2.1.1. **Unassessed:** 1.4.3. ## 🌍 DEIB Assessment #### Cultural breadth **Narrow by design.** Two US-headquartered multinationals, one US lab, English only, corporate register. "Billions of consumers around the world" is the sole gesture outward and it's a scale claim, not a perspective. #### Inclusive language **Neutral.** No gendered, evaluative, or exclusionary language. "Responsible stewards" is the closest thing to a values term and it's conventional. #### Neurodiversity support **Weak.** Abstract nouns, fast cuts between unidentified speakers, no signposting, no summary, no visual structure for a passage that's a five-item list. A viewer who processes literally has to reconstruct who's speaking and what "signals," "findings," and "safeguards" mean from context the clip never gives. #### Representation **Not assessable from the transcript.** Speaker demographics aren't captured. Institutionally: two customer voices, both from very large incumbents; no small-business, public-sector, or non-US perspective on what "safeguards" should mean. #### Psychological safety **Not applicable in the interpersonal sense.** In the systemic sense, the architecture described — machine-only review, no human reads your content — is itself a psychological-safety claim for end users whose payment or ride data is in scope. It's the one part of the clip that speaks to the people whose data this is, and it does so indirectly, via the customer's executive. ## ✅ Logical Validity **Strong reasoning.** - *Architecture over policy.* The claim that controls built into the system are stronger than controls written into policy is correct and the clip states it cleanly. It's the same ranking as "a hook beats a README." - *Findings-only output as privacy boundary.* Limiting what leaves the review to a defined schema is a sound minimization principle; the reasoning is implicit but valid. **Fallacies.** - **Appeal to authority / brand credentialing** — the reel's spine. Visa's 130 years and 15,000 banks establish that Visa is trusted, not that the product is trustworthy. Located throughout the Visa segment. *Fix:* pair each credential with a verifiable property (which the architecture passage partly does). `#fallacy/appeal-to-authority` - **Testimonial as evidence** — customers with a commercial relationship attesting to the vendor's product. Standard for the genre, still a fallacy if consumed as proof. *Fix:* independent attestation or published documentation. `#fallacy/anecdotal` - **Unstated assumption** — "Visa cannot do it ourselves" presumes the capability gap; no reason is given why a company of Visa's resources couldn't build machine-only review with findings-only output, which is a pattern its own DLP tooling already uses. *Fix:* say what specifically requires a frontier lab (the model, presumably) rather than implying the whole architecture does. `#fallacy/false-dilemma` - No DARVO, no straw man, no ad hominem. The reel attacks no one. **Tone of voice.** Executive-confident, values-forward, low on specifics except in one passage. "I don't intend to spend it carelessly" is the one line with a person in it. ## 🛡️ DARVO Analysis **Not applicable to this source.** No accusation exists to deny, no critic to attack, no victim/offender positions to reverse. The reel is promotional, not defensive. The nearest thing to a defensive move is pre-emptive — "now that the stakes are higher" acknowledges a risk environment without naming any critic or incident — and pre-emption isn't DARVO. ## 🎯 Conclusion **Bottom line.** Two sentences of architecture inside two minutes of credentials. Keep the architecture; ignore the credentials. **What holds up.** Five stated properties — customer-cloud residency, continuous signal access, customer-controlled egress, machine-only review, findings-only output — each of which is checkable against documentation if documentation exists. "Architectural level, not policy level" as a compact statement of the enforcement ladder. **What doesn't.** Every voice has a commercial stake. No definitions. No speaker names in the transcript. No statement of where inference runs, which is the first thing a security team asks. "Visa cannot do it ourselves" asserts a gap it doesn't explain. **Next steps.** 1. Find the product docs for "Enterprise Frontier Safeguards" and reconcile the five spoken properties against them; log any gap. 2. Cross-link the "architecture not policy" line to Courtney's own enforcement-ladder writing — same claim, industry-scale instance — rather than creating a new note. 3. Do not cite this clip as evidence the safeguards work. Cite it, if at all, as evidence of what Anthropic is *claiming* to enterprise buyers in September 2026. ## 🔄 Refactor Appendix — Building Enterprise Frontier Safeguards with our customers ### 📚 Literature Atoms (H3) #### [H1] Controls at the architectural level are not offered as a choice > [!quote] ID:rw_01m1fjjycea6e4mpa18kfc0jnb > "We were able to work together on new security and privacy capabilities at the architectural level, not just policy level." A policy is a rule a person can skip; an architecture is a rule that isn't presented for skipping. The Visa speaker names the distinction in one clause. Same ranking as write-gate over written instruction. Tags: `#atom/literature/h1` #### [H1] A review that emits findings, not content, keeps the reviewer from becoming a data recipient > [!quote] ID:rw_01m1fjjycea6e4mpa18kfc0jnb > "The review is machine-only. What comes out is intentionally limited to defined findings, not customer content." The privacy pattern: the boundary is defined by the output schema. Linters, DLP scanners, and moderation APIs already work this way; here it's applied to frontier-model review over regulated data. Tags: `#atom/literature/h1` #### [H3] Trust is framed as a depletable asset, not a credential > [!quote] ID:rw_01m1fjjycea6e4mpa18kfc0jnb > "That trust took decades to earn, and I don't intend to spend it carelessly now that the stakes are higher." The one line in the reel with a person and a risk in it. Trust as something you *spend* reframes AI adoption as a withdrawal, not a feature. Tags: `#atom/literature/h3` ### 🧠 Permanent Candidates (H4) #### [P1] The privacy boundary of a review system is its output schema What a monitoring or review process may emit defines what it can leak. Fix the schema of findings first — categories, severities, pointers — and the reviewer, human or machine, is structurally prevented from becoming a recipient of the content it inspected. Prompts, models, and policies are implementation details behind that contract. The pattern is old (linters, DLP, moderation APIs); the Visa passage applies it to frontier-model review over payment data. Links: [[An Upgrade Review Diffs the Defaults Not the API]] · [[Data minimization]] · [[AI Agent Safety MOC]] Tags: `#atom/zettel/z1` #### [P2] A rule enforced by architecture is one a person was never offered the chance to break Policies are enforced probabilistically — someone has to read, remember, and choose. Architecture removes the choice. The same ranking appears in tooling (pre-commit hook over CONTRIBUTING.md), in the deep-read enforcement ladder (write gate over instruction), and here in enterprise AI controls ("architectural level, not just policy level"). When a rule keeps failing as a policy, the fix is to push it down a rung, not to write it louder. *Check for an existing Permanent on the enforcement ladder before creating — this may already be captured.* Links: [[Default-On AI Activation]] · [[Alarm fatigue before adding alarms]] · [[Enforcement ladder]] Tags: `#atom/zettel/z1` ### ✏️ Fleeting Queue - The clip says logs stay in the customer's cloud. It never says the model does. The omission looks deliberate. `#atom/fleeting/f1` - "Visa cannot do it ourselves" — Visa runs DLP tooling with findings-only output already. What specifically needs a frontier lab? The model, presumably; the architecture, no. `#atom/fleeting/f1` - Two multi-speaker clips in this batch with no speaker labels in captions. Anthropic's own channel doesn't label speakers. Note for the accessibility column. `#atom/fleeting/f1` - Trust as a *spendable* asset — Visa's line — pairs with the deconstruction material on institutional trust. Different domain, same mechanics. `#atom/fleeting/f1` ### → Things 3 (max 2 actions) 1. **Locate Anthropic's documentation for "Enterprise Frontier Safeguards"** and check the five spoken properties (residency, observability, egress, machine-only review, findings-only output) against it. `#action/h2` 2. **Grep the vault for an existing enforcement-ladder Permanent** before writing P2 — the deep-read tooling doc already carries the argument. `#action/h3` <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Building Enterprise Frontier Safeguards With Our Customers.md|Building Enterprise Frontier Safeguards With Our Customers by Claude (Anthropic)]] ## Highlights - We were able to work together on new security and privacy capabilities at the architectural level, not just policy level. We have the ability to store in our cloud with our control, and that we have continuous access to the signals. The logs are under our control. ([View Highlight](https://read.readwise.io/read/01m1fsvyyy0352ak1hxbmj3d45)) - Note: #atom/literature/h1 — Architecture over policy: a control built into the system isn't offered as a choice. Same ranking as write-gate over written instruction. - Tags: atom/literature/h1 - They don't go anywhere else unless we want them to. The review is machine-only. What comes out is intentionally limited to defined findings, not customer content. Visa cannot do it ourselves. We need the frontier labs like Anthropic helping us to provide the ingredients to help us to bring AI into every aspect of our cyber defense. Those safeguards actually allow us to apply AI in parts of the business ([View Highlight](https://read.readwise.io/read/01m1fsw3cbtw0xvdas8fkp0vvh)) - Note: #atom/literature/h1 — Findings-only output as the privacy boundary: machine-only review that emits a defined schema, never the content it inspected. Note the omission — logs stay in the customer cloud; nothing says the model does. - Tags: atom/literature/h1