https://montanalabs.ai/news/ Analysis, field notes, and editorial coverage on AI products, workflow automation, frontend systems, and platform delivery. en-US Tue, 14 Jul 2026 00:00:00 GMT https://montanalabs.ai/news/how-google-deepmind-rebuilt-pel-s-unfilmed-1959-goal-from-archives-and-stunt-footage/ https://montanalabs.ai/news/how-google-deepmind-rebuilt-pel-s-unfilmed-1959-goal-from-archives-and-stunt-footage/ Tue, 14 Jul 2026 00:00:00 GMT AI Products Google's reconstruction of a goal that was never recorded is less a demo of generative video than a case study in constraining it — with 3,600 archival images, motion-captured stunt performances, and a layered post-production pipeline built to enforce historical accuracy. https://montanalabs.ai/news/dnp-put-chatgpt-enterprise-in-front-of-ten-departments-and-treated-the-chat-window-as-the-interface/ https://montanalabs.ai/news/dnp-put-chatgpt-enterprise-in-front-of-ten-departments-and-treated-the-chat-window-as-the-interface/ Mon, 13 Jul 2026 00:00:00 GMT Frontend Dai Nippon Printing's ChatGPT Enterprise rollout is less a model story than an interface story: the chat window became the single frontend through which non-programmers ran code, IP teams searched patents, and paper manuals became queryable knowledge bases. https://montanalabs.ai/news/adventhealth-deploys-chatgpt-across-nine-states-by-treating-adoption-as-the-product/ https://montanalabs.ai/news/adventhealth-deploys-chatgpt-across-nine-states-by-treating-adoption-as-the-product/ Mon, 13 Jul 2026 00:00:00 GMT Frontend AdventHealth's ChatGPT rollout is a lesson in what happens when the user-facing product is a blank text box. The organization reports an 80% reduction in time on administrative tasks — but the interesting work was everything wrapped around the interface: peer groups, a usage metric, and workflow redesign anchored to EHR timestamps. https://montanalabs.ai/news/openai-reframes-adoption-as-a-capability-overhang-problem/ https://montanalabs.ai/news/openai-reframes-adoption-as-a-capability-overhang-problem/ Mon, 13 Jul 2026 00:00:00 GMT Automation OpenAI's 'AI for self empowerment' post introduces 'capability overhang' as its organizing idea and ties adoption to compute consumption, ad-supported free access, and its own economic benchmarks. https://montanalabs.ai/news/expedia-s-image-selection-automation-is-the-concrete-piece-behind-its-ai-marketing-story/ https://montanalabs.ai/news/expedia-s-image-selection-automation-is-the-concrete-piece-behind-its-ai-marketing-story/ Mon, 13 Jul 2026 00:00:00 GMT AI Products Expedia Group is using OpenAI APIs to pick and present the right images across its travel sites, and ChatGPT across its marketing team. We separate the deployed work from the forecasting. https://montanalabs.ai/news/ap-uses-codex-to-build-behaving-payment-prototypes-not-just-clickable-screens/ https://montanalabs.ai/news/ap-uses-codex-to-build-behaving-payment-prototypes-not-just-clickable-screens/ Mon, 13 Jul 2026 00:00:00 GMT Frontend AP+ reports building working payment-flow simulations in a day with Codex, moving from screens that show what an experience looks like to prototypes that behave closer to real systems. https://montanalabs.ai/news/autoscout24-puts-codex-in-front-of-1-000-builders-to-speed-marketplace-delivery/ https://montanalabs.ai/news/autoscout24-puts-codex-in-front-of-1-000-builders-to-speed-marketplace-delivery/ Mon, 13 Jul 2026 00:00:00 GMT Frontend AutoScout24 Group reports cutting select project timelines from 2–3 weeks to 2–3 days after a three-month Codex evaluation, while extending prototyping to non-engineers. The interesting part is the structure it built to get there. https://montanalabs.ai/news/bbva-s-ai-rollout-put-the-gpt-builder-in-front-of-100-000-bank-employees/ https://montanalabs.ai/news/bbva-s-ai-rollout-put-the-gpt-builder-in-front-of-100-000-bank-employees/ Mon, 13 Jul 2026 00:00:00 GMT Frontend BBVA moved from 3,000 seats in 2024 to more than 100,000 employees on ChatGPT Enterprise, with staff building over 20,000 custom GPTs. The interesting part is where the interface lives: not a single flagship app, but thousands of small, department-owned front ends. https://montanalabs.ai/news/choco-built-an-inference-layer-to-encode-order-desk-knowledge-not-just-transcribe-orders/ https://montanalabs.ai/news/choco-built-an-inference-layer-to-encode-order-desk-knowledge-not-just-transcribe-orders/ Mon, 13 Jul 2026 00:00:00 GMT Frontend Choco's OrderAgent and VoiceAgent process 8.8M+ orders annually across 21,000 distributors. The interesting engineering isn't the multimodal input — it's the dynamic in-context learning that resolves customer-specific ambiguity, plus a confidence-gated Autopilot that keeps humans on edge cases. https://montanalabs.ai/news/cisco-built-the-majority-of-its-ai-defense-product-with-codex-writing-the-code/ https://montanalabs.ai/news/cisco-built-the-majority-of-its-ai-defense-product-with-codex-writing-the-code/ Mon, 13 Jul 2026 00:00:00 GMT Automation Cisco reports 95%+ of new AI features written by Codex and 1,500+ engineering hours saved per month. The specifics of where those gains came from matter more than the headline numbers. https://montanalabs.ai/news/commonwealth-bank-standardizes-on-chatgpt-enterprise-as-the-shared-surface-for-50-000-employees/ https://montanalabs.ai/news/commonwealth-bank-standardizes-on-chatgpt-enterprise-as-the-shared-surface-for-50-000-employees/ Mon, 13 Jul 2026 00:00:00 GMT Automation Commonwealth Bank of Australia is deploying ChatGPT Enterprise to nearly 50,000 employees, choosing a consistent, familiar product surface over bespoke internal tooling — and naming customer service and fraud response as where agents come next. https://montanalabs.ai/news/consensus-rebuilt-its-science-search-as-a-four-agent-workflow-on-gpt-5-and-the-responses-api/ https://montanalabs.ai/news/consensus-rebuilt-its-science-search-as-a-four-agent-workflow-on-gpt-5-and-the-responses-api/ Mon, 13 Jul 2026 00:00:00 GMT Frontend Consensus split its research workflow into four scoped agents on GPT-5 and migrated to the Responses API for routing, while keeping a fast, conversational front end aimed directly at the people doing the research. https://montanalabs.ai/news/cyberagent-embeds-codex-into-design-and-mockup-work-not-just-code-generation/ https://montanalabs.ai/news/cyberagent-embeds-codex-into-design-and-mockup-work-not-just-code-generation/ Mon, 13 Jul 2026 00:00:00 GMT Frontend CyberAgent's account of ChatGPT Enterprise and Codex adoption is notable less for headline usage numbers than for where the tools landed: upstream design decisions, mockups, and the internal systems that sustain adoption itself. https://montanalabs.ai/news/databricks-puts-gpt-5-5-in-charge-of-enterprise-document-agents/ https://montanalabs.ai/news/databricks-puts-gpt-5-5-in-charge-of-enterprise-document-agents/ Mon, 13 Jul 2026 00:00:00 GMT Frontend Databricks reports GPT-5.5 crossed 50% on its OfficeQA Pro benchmark and cut errors 46% versus GPT-5.4, with the biggest gains in parsing older documents — the exact failure point that breaks production agents. https://montanalabs.ai/news/datadog-validated-codex-against-its-own-past-incidents-before-rolling-it-out/ https://montanalabs.ai/news/datadog-validated-codex-against-its-own-past-incidents-before-rolling-it-out/ Mon, 13 Jul 2026 00:00:00 GMT Automation Datadog didn't accept Codex on demo footing. Its engineers rebuilt pull requests that had caused real incidents, ran Codex against them, and asked the people who owned those incidents whether the feedback would have changed anything. The 22% hit rate is the number worth studying. https://montanalabs.ai/news/doppel-automates-phishing-takedowns-with-a-five-stage-gpt-5-and-rft-pipeline/ https://montanalabs.ai/news/doppel-automates-phishing-takedowns-with-a-five-stage-gpt-5-and-rft-pipeline/ Mon, 13 Jul 2026 00:00:00 GMT Platform Doppel's account of its rollout is unusually specific about where AI sits in its pipeline and what problem reinforcement fine-tuning actually solved: not accuracy, but the inconsistency of human analysts judging the same ambiguous threat differently. https://montanalabs.ai/news/eneos-materials-built-over-1-000-custom-gpts-and-put-chatgpt-enterprise-in-front-of-every-employee/ https://montanalabs.ai/news/eneos-materials-built-over-1-000-custom-gpts-and-put-chatgpt-enterprise-in-front-of-every-employee/ Mon, 13 Jul 2026 00:00:00 GMT AI Products ENEOS Materials adopted ChatGPT Enterprise company-wide, and its results hinge less on model capability than on who is allowed to use the interface. The most telling detail is an HR staffer with no coding experience building a working data-aggregation tool by talking to the model in Japanese. https://montanalabs.ai/news/hibob-prototypes-customer-features-inside-chatgpt-before-shipping-them-via-api/ https://montanalabs.ai/news/hibob-prototypes-customer-features-inside-chatgpt-before-shipping-them-via-api/ Mon, 13 Jul 2026 00:00:00 GMT Frontend HiBob's account describes an unusual loop: the same custom GPTs that employees use internally double as prototypes for features that later ship in the Bob platform through OpenAI's API. The frontend detail worth studying is where the conversational interface lives at each stage. https://montanalabs.ai/news/axios-built-a-custom-gpt-to-run-local-news-in-one-reporter-cities/ https://montanalabs.ai/news/axios-built-a-custom-gpt-to-run-local-news-in-one-reporter-cities/ Mon, 13 Jul 2026 00:00:00 GMT Frontend Axios isn't automating journalism wholesale. It's stripping production tasks off reporters so a single person can launch a city, and OpenAI is paying to prove the model in Boulder, Huntsville, Pittsburgh, and Kansas City. https://montanalabs.ai/news/balyasny-embedded-openai-models-into-180-investment-teams-daily-workflows-and-95-now-use-it/ https://montanalabs.ai/news/balyasny-embedded-openai-models-into-180-investment-teams-daily-workflows-and-95-now-use-it/ Mon, 13 Jul 2026 00:00:00 GMT Automation Balyasny Asset Management reached 95% adoption across its investment teams by designing the surface where analysts meet the model: scoped tools, traceable reasoning, federated deployment, and feedback captured in real time. https://montanalabs.ai/news/bbva-turned-20-000-custom-gpts-into-its-ai-front-end/ https://montanalabs.ai/news/bbva-turned-20-000-custom-gpts-into-its-ai-front-end/ Mon, 13 Jul 2026 00:00:00 GMT Frontend BBVA reports 83% weekly active usage and 20,000+ Custom GPTs across the bank. The interesting part isn't the headline savings — it's who built the interfaces employees actually touch. https://montanalabs.ai/news/braintrust-moved-half-its-engineering-team-to-codex-in-a-month-by-collapsing-the-feature-request-backlog/ https://montanalabs.ai/news/braintrust-moved-half-its-engineering-team-to-codex-in-a-month-by-collapsing-the-feature-request-backlog/ Mon, 13 Jul 2026 00:00:00 GMT Frontend Braintrust reports that half its team adopted Codex with GPT-5.5 in one month. The interesting detail isn't the coding speed itself but what it removed: the backlog step between a customer request and a working preview. https://montanalabs.ai/news/chime-s-marketing-team-built-a-custom-gpt-on-its-own-best-performing-content/ https://montanalabs.ai/news/chime-s-marketing-team-built-a-custom-gpt-on-its-own-best-performing-content/ Mon, 13 Jul 2026 00:00:00 GMT Automation The most concrete detail in Chime's account isn't the agentic-marketing rhetoric — it's that its personalization and bidding models underperformed until the inputs were fixed, and that off-the-shelf models were replaced with one trained on Chime's own copy. https://montanalabs.ai/news/deutsche-telekom-s-bet-that-voice-networks-become-the-ai-interface/ https://montanalabs.ai/news/deutsche-telekom-s-bet-that-voice-networks-become-the-ai-interface/ Mon, 13 Jul 2026 00:00:00 GMT Platform Deutsche Telekom frames its OpenAI work as an operating-model redesign, not a tool rollout. The most concrete signal is its plan to put live translation, in-call assistants, and post-call summaries directly into the voice network — no new app required. https://montanalabs.ai/news/endava-routes-codex-to-the-front-of-the-delivery-pipeline-not-just-the-code-editor/ https://montanalabs.ai/news/endava-routes-codex-to-the-front-of-the-delivery-pipeline-not-just-the-code-editor/ Mon, 13 Jul 2026 00:00:00 GMT Automation Endava's most interesting use of Codex isn't autocomplete. It's generating requirements specs from meeting transcripts and drawing architecture diagrams live in client sessions — pushing the agent to the earliest, most human-facing stages of a project. https://montanalabs.ai/news/endava-s-davaflow-puts-openai-in-every-stage-of-software-delivery/ https://montanalabs.ai/news/endava-s-davaflow-puts-openai-in-every-stage-of-software-delivery/ Mon, 13 Jul 2026 00:00:00 GMT Frontend Endava embedded ChatGPT Enterprise and Codex across an 11,000-person workforce, but the detail worth studying is what happened when a pricing discussion produced a single-page app instead of a spreadsheet. https://montanalabs.ai/news/gpt-5-as-a-proof-exploration-tool-what-ernest-ryu-s-nag-result-actually-shows/ https://montanalabs.ai/news/gpt-5-as-a-proof-exploration-tool-what-ernest-ryu-s-nag-result-actually-shows/ Mon, 13 Jul 2026 00:00:00 GMT Frontend OpenAI's account of Ernest Ryu's work on the Nesterov Accelerated Gradient problem is less a story about a model solving math than about a specific human workflow: fast idea generation, aggressive discarding, and relentless verification in fresh chats. https://montanalabs.ai/news/indeed-puts-two-ai-agents-in-front-of-job-seekers-and-recruiters/ https://montanalabs.ai/news/indeed-puts-two-ai-agents-in-front-of-job-seekers-and-recruiters/ Mon, 13 Jul 2026 00:00:00 GMT AI Products Indeed's OpenAI interview is mostly about the interface layer: turning a job board into two conversational agents while insisting the recommendation surface show its reasoning and keep the human as the decision-maker. https://montanalabs.ai/news/hygh-built-a-camera-to-billboard-preview-tool-on-openai-s-apis/ https://montanalabs.ai/news/hygh-built-a-camera-to-billboard-preview-tool-on-openai-s-apis/ Mon, 13 Jul 2026 00:00:00 GMT Frontend HYGH's most concrete result isn't the 5.5 hours saved per employee — it's a shipped API-powered interface that renders a client's product onto its network of 4,000+ screens in real time. https://montanalabs.ai/news/intercom-s-model-agnostic-frontend-how-fin-ships-new-modalities-without-a-rewrite/ https://montanalabs.ai/news/intercom-s-model-agnostic-frontend-how-fin-ships-new-modalities-without-a-rewrite/ Mon, 13 Jul 2026 00:00:00 GMT Frontend Intercom's Fin resolves millions of queries a month across chat, email, and voice. What stands out in OpenAI's writeup is not the model choice but the frontend architecture that lets Intercom swap models in days and add whole new interaction modes without reengineering. https://montanalabs.ai/news/coderabbit-routes-code-review-across-three-openai-models-to-attack-the-shipping-bottleneck/ https://montanalabs.ai/news/coderabbit-routes-code-review-across-three-openai-models-to-attack-the-shipping-bottleneck/ Mon, 13 Jul 2026 00:00:00 GMT Frontend CodeRabbit's pitch inverts the usual AI coding story: the constraint isn't writing code, it's reviewing it. Their multi-model system and pre-merge timing choice are the interesting engineering decisions here. https://montanalabs.ai/news/meta-s-5gw-louisiana-expansion-is-announced-through-teacher-bonuses-not-teraflops/ https://montanalabs.ai/news/meta-s-5gw-louisiana-expansion-is-announced-through-teacher-bonuses-not-teraflops/ Mon, 13 Jul 2026 00:00:00 GMT Platform Meta's Richland Parish announcement leads with $50,000 teacher bonuses and taco shops, not GPUs. Reading the compute, contracting, and energy details underneath the community narrative. https://montanalabs.ai/news/bny-put-ai-governance-inside-the-interface-not-around-it/ https://montanalabs.ai/news/bny-put-ai-governance-inside-the-interface-not-around-it/ Mon, 13 Jul 2026 00:00:00 GMT Frontend BNY's internal platform Eliza makes prompting, model selection, agent building, and sharing happen inside a single governed environment. The interesting engineering choice is where the controls live: in the tooling employees touch, not in a separate review step bolted on afterward. https://montanalabs.ai/news/openai-packages-chatgpt-into-a-school-domain-workspace-for-u-s-k-12-teachers/ https://montanalabs.ai/news/openai-packages-chatgpt-into-a-school-domain-workspace-for-u-s-k-12-teachers/ Thu, 09 Jul 2026 00:00:00 GMT Frontend ChatGPT for Teachers isn't a new model — it's a repackaging of the existing product into an organization-scoped frontend, complete with SheerID verification, role-based access controls, and prompt examples surfaced under the composer. https://montanalabs.ai/news/openai-s-gpt-5-codex-addendum-an-agentic-coder-trained-to-match-pr-taste-and-run-its-own-tests/ https://montanalabs.ai/news/openai-s-gpt-5-codex-addendum-an-agentic-coder-trained-to-match-pr-taste-and-run-its-own-tests/ Thu, 09 Jul 2026 00:00:00 GMT Frontend OpenAI's addendum for GPT-5-Codex describes a model trained by reinforcement learning to mirror human code style, follow instructions precisely, and loop on tests until they pass — plus sandboxing and configurable network access for when it runs as an agent. https://montanalabs.ai/news/openai-stacks-three-provenance-layers-c2pa-conformance-synthid-watermarks-and-a-public-verifier/ https://montanalabs.ai/news/openai-stacks-three-provenance-layers-c2pa-conformance-synthid-watermarks-and-a-public-verifier/ Thu, 09 Jul 2026 00:00:00 GMT AI Products OpenAI's provenance update pairs C2PA conformance with SynthID watermarking and a cautious public verification tool—an admission that no single signal survives the way content actually travels online. https://montanalabs.ai/news/openai-reports-gpt-5-2-pro-solved-an-open-statistics-problem-posed-at-colt-2019/ https://montanalabs.ai/news/openai-reports-gpt-5-2-pro-solved-an-open-statistics-problem-posed-at-colt-2019/ Thu, 09 Jul 2026 00:00:00 GMT Frontend OpenAI's GPT-5.2 post pairs headline benchmark numbers with a single documented case study: a model asked to resolve an open problem in learning theory directly, then checked by external experts. The workflow it describes is more instructive than the scores. https://montanalabs.ai/news/openai-splits-realtime-voice-into-three-models-reasoning-translation-and-transcription/ https://montanalabs.ai/news/openai-splits-realtime-voice-into-three-models-reasoning-translation-and-transcription/ Thu, 09 Jul 2026 00:00:00 GMT Platform OpenAI's new audio lineup treats a voice agent as three separable jobs—reason, translate, transcribe—and gives developers explicit controls over latency, context length, and tool transparency. https://montanalabs.ai/news/openai-proposes-an-international-youth-ai-safety-institute-ahead-of-the-g7-in-vian/ https://montanalabs.ai/news/openai-proposes-an-international-youth-ai-safety-institute-ahead-of-the-g7-in-vian/ Thu, 09 Jul 2026 00:00:00 GMT AI Products OpenAI's call for a dedicated youth safety institute is timed to the G7 Leaders' Summit, but the operational substance sits in the principles it wants governments to adopt: knowing when a user is a minor, defaulting to protection when age is uncertain, and making audits interoperable across borders. https://montanalabs.ai/news/gpt-5-1-codex-max-and-the-cost-curve-of-generated-frontends/ https://montanalabs.ai/news/gpt-5-1-codex-max-and-the-cost-curve-of-generated-frontends/ Thu, 09 Jul 2026 00:00:00 GMT Frontend OpenAI says GPT-5.1-Codex-Max produces frontend designs comparable to its predecessor at lower token cost, and can run for hours through a technique it calls compaction. We read the frontend and long-horizon claims against what the source actually documents. https://montanalabs.ai/news/openai-puts-checkout-inside-the-chat-window-with-instant-checkout-and-the-agentic-commerce-protocol/ https://montanalabs.ai/news/openai-puts-checkout-inside-the-chat-window-with-instant-checkout-and-the-agentic-commerce-protocol/ Thu, 09 Jul 2026 00:00:00 GMT Frontend OpenAI's Instant Checkout turns a chat answer into a transaction surface. The interesting details are in how the flow keeps the user confirming each step while the merchant stays the merchant of record. https://montanalabs.ai/news/openai-s-chatgpt-agent-operates-through-a-rendered-browser-not-an-api/ https://montanalabs.ai/news/openai-s-chatgpt-agent-operates-through-a-rendered-browser-not-an-api/ Thu, 09 Jul 2026 00:00:00 GMT Frontend OpenAI's ChatGPT agent system card describes a model that browses in a remote visual environment, runs a terminal with limited network access, and reaches into Google Drive through Connectors. For teams building web frontends, the notable detail is that the agent perceives applications the way a person does — through what the page renders. https://montanalabs.ai/news/openai-folds-sites-a-built-in-browser-and-computer-use-into-chatgpt-work/ https://montanalabs.ai/news/openai-folds-sites-a-built-in-browser-and-computer-use-into-chatgpt-work/ Thu, 09 Jul 2026 00:00:00 GMT Frontend With Sites in public beta, a desktop browser, and the Atlas standalone browser being sunset, OpenAI is moving generated interfaces and web interaction into a single ChatGPT surface. https://montanalabs.ai/news/openai-adds-an-in-app-browser-and-computer-use-to-codex-aimed-squarely-at-frontend-iteration/ https://montanalabs.ai/news/openai-adds-an-in-app-browser-and-computer-use-to-codex-aimed-squarely-at-frontend-iteration/ Thu, 09 Jul 2026 00:00:00 GMT Frontend Codex now includes an in-app browser you can comment on, background computer use on macOS, and gpt-image-1.5 image generation. The through-line is closing the visual feedback loop that text-only coding agents never had. https://montanalabs.ai/news/openai-puts-a-hosted-frontend-behind-codex-with-sites-annotations-and-a-product-design-plugin/ https://montanalabs.ai/news/openai-puts-a-hosted-frontend-behind-codex-with-sites-annotations-and-a-product-design-plugin/ Thu, 09 Jul 2026 00:00:00 GMT Frontend Codex's new Sites, annotations, and product design plugin move it from writing code to producing and refining hosted frontend artifacts—competing directly with the tools it names as partners. https://montanalabs.ai/news/openai-ships-codex-ga-with-a-typescript-first-sdk-and-slack-agent/ https://montanalabs.ai/news/openai-ships-codex-ga-with-a-typescript-first-sdk-and-slack-agent/ Thu, 09 Jul 2026 00:00:00 GMT Frontend Codex's GA bundles a Slack integration, a TypeScript SDK, and admin controls. The detail worth noting is where the agent now runs and which language got it first. https://montanalabs.ai/news/openai-splits-codex-billing-from-chatgpt-seats-with-token-based-pricing/ https://montanalabs.ai/news/openai-splits-codex-billing-from-chatgpt-seats-with-token-based-pricing/ Thu, 09 Jul 2026 00:00:00 GMT Frontend OpenAI decoupled Codex access from fixed ChatGPT seat fees and moved it to token consumption — then, within weeks, closed the new pay-as-you-go path for Business plans while leaving Enterprise and existing seats intact. https://montanalabs.ai/news/openai-ships-codex-security-an-application-security-agent-built-on-a-per-project-threat-model/ https://montanalabs.ai/news/openai-ships-codex-security-an-application-security-agent-built-on-a-per-project-threat-model/ Thu, 09 Jul 2026 00:00:00 GMT Frontend Codex Security grounds vulnerability discovery in a project-specific threat model and sandboxed validation, targeting the exact class of auth, SSRF, and session bugs that surface when web applications ship quickly. https://montanalabs.ai/news/openai-loses-its-whatsapp-channel-and-has-to-migrate-50-million-users-by-hand/ https://montanalabs.ai/news/openai-loses-its-whatsapp-channel-and-has-to-migrate-50-million-users-by-hand/ Thu, 09 Jul 2026 00:00:00 GMT Frontend OpenAI is shutting down ChatGPT on WhatsApp after a policy change from the platform, and the migration reveals how fragile a borrowed distribution channel really is when you can't export the data users created there. https://montanalabs.ai/news/how-sora-bakes-safety-into-the-product-surface-not-just-the-model/ https://montanalabs.ai/news/how-sora-bakes-safety-into-the-product-surface-not-just-the-model/ Thu, 09 Jul 2026 00:00:00 GMT Frontend OpenAI's Sora safety post reads as much like a product spec as a policy statement. The protections it describes — watermarks, consent attestations, character permissions, teen scroll limits — are things a user encounters in the interface, which makes the front end the enforcement layer.