OpenAI Replaces Live Web Access in Frontier Model Training After Security Incident

·9 views

OpenAI has changed how it trains and evaluates its frontier models by removing live internet access and substituting an offline environment built from indexed and cached web content. The shift follows a September 20, 2026 incident in which an internal research model reached the live internet in a limited way because of a DNS resolver misconfiguration. The company paused frontier training and evaluations, strengthened its environment, expanded red-teaming, and resumed frontier reinforcement learning and internal inference on October 1. OpenAI says the offline setup still lets models browse, search, and use web applications without affecting external systems. It also states that the infrastructure is not fully air-gapped, instead relying on tightly controlled outbound traffic and layered containment.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Free FMCSA census query surfaces newly registered trucking companies

A developer shared a method for pulling newly registered trucking companies from the FMCSA carrier census, which is published as open data on data.transportation.gov. The approach uses a SoQL query over plain HTTP, filtering active records by registration date and state, and requires no API key for light use. The author notes that the raw dataset lacks built-in deduplication, operating authority status, and normalized fields, which must be handled separately. To address these gaps, the author built a paid Apify actor that joins authority data, applies filters, avoids repeat deliveries, and exports results in JSON, CSV, or Excel. The article notes the author's bias, since they built the tool they promote.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Expo Router navigation slows past 40 nested routes, with fixes for layout re-renders and focus

A team migrating a multi-tenant dashboard from React Navigation to Expo Router found a performance drop at around 42 nested routes, with animations falling from 60fps to 38fps and memory rising 140MB on Android mid-range devices. Benchmarks on a Pixel 5 and iPhone 12 using Expo SDK 52 and React Native 0.76 showed navigation time and frame drops rising sharply as route depth increased. The team attributed the problem to re-renders of the layout tree on navigation events, compounded by runtime TailwindCSS class recalculation in nested layouts. Replacing dynamic class props with static StyleSheet references cut navigation time at 50 routes from 187ms to 94ms and reduced frame drops per transition from 9 to 3. The team also found that modal routes overlaying tab navigators broke screen reader focus, which they addressed by wrapping the modal in a container with accessibilityViewIsModal.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Plex Media Server flaw CVE-2026-96655 can be checked in logs for prior exploitation

Patching Plex Media Server to 1.43.3.10861 closes the SSRF in CVE-2026-96655, but it does not reveal whether the flaw was already abused. The bug allows an authenticated user to make the server fetch an attacker-chosen target via the universal transcoder endpoint. Investigators are advised to compare endpoint activity against a baseline of normal household use and to look for internal network connections from the Plex host that it would not normally make. Failed transcode sessions can serve as a weak triage signal but are not conclusive evidence. If compromise is suspected, logs and connection records should be preserved, the involved account disabled, and any exposed credentials rotated before patching.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Video generation models still face a wide gap between demos and production use

Video models can produce impressive short clips in marketing demos, but production pipelines run into latency problems, weak frame-to-frame consistency, limited camera control, and mandatory brand-safety review. Treating video generation as simply a longer form of image generation tends to underestimate costs and iteration cycles. Teams that successfully ship multimodal features usually begin with narrower tasks such as understanding clips, captioning meetings, or extracting frames for search. Assisted editing, storyboards, and animatics are also more practical than generating final pixels for regulated brands. The article notes that legal and brand review often takes more time than GPU inference, and it recommends tools with controllable parameters and evaluation methods for temporal artifacts.

0 comments Read more at DEV Community