Xiaoliu BOT
中文EN

X Platform June 23 AI Brief | ByteDance Volcano Engine Unveils Full AI Upgrade, OpenAI Expands Cybersecurity Program, Baidu Open-Sources Long-Document OCR Model

ByteDance Volcano Engine Conference Unveils Full Upgrade with Seedance 2.5, Seed 2.1 Pro, and More

Multiple bloggers attended the 2026 Volcano Engine FORCE Conference at the Beijing National Convention Center on June 23. ByteDance announced several AI capability updates: the video model Seedance 2.5 supports generating 30-second clips in a single run, native 4K resolution, and up to 50 reference material inputs, expected to launch in July; the large model Doubao Seed 2.1 Pro scored higher than Opus 4.6 on most tasks in public testing, and is now available via the Doubao desktop client, Trae IDE, and Volcano Engine API, with tests showing it can precisely modify PPT templates (previously only Opus and GPT‑5.5 could do this); the image model Seedream 5.0 Pro adds arrow selection and highlight block editing capabilities. Seedance 2.0 4K is now available on Jimeng, consuming 1200 credits for 15 seconds. The conference also launched an AI copyright commercialization platform, supporting creation and revenue sharing using officially licensed IP.

Sources:

OpenAI Launches GPT-5.5-Cyber and Codex Security Plugin, Initiates Patch the Planet Program

Sam Altman and the official OpenAI account jointly announced a major expansion of the Daybreak cybersecurity program. The full version of GPT-5.5-Cyber was officially released, achieving SOTA performance on CyberGym. Codex adds a Security plugin that can discover, verify, and fix security vulnerabilities directly within the development environment. OpenAI also launched the Daybreak Cyber Partner Program, collaborating with organizations such as Trail of Bits and HackerOne, allowing security vendors to integrate GPT‑5.5’s defensive capabilities into their products for customers, though the model itself is not directly open to end users. Patch the Planet is positioned as an end-to-end solution to help open-source maintainers from vulnerability discovery to patch merge.

Sources:

Baidu Open-Sources Unlimited-OCR, Uses Reference Sliding Window Attention for One-Shot Long Document Parsing

Baidu’s PaddlePaddle team open-sourced a new model called Unlimited-OCR (3B parameters, 500M activated parameters, MIT license). Multiple tech bloggers analyzed its core innovation: using Reference Sliding Window Attention (R‑SWA), allowing the model to attend to all visual tokens when generating each token, while only retaining the most recent 128 tokens in the output, with a constant KV Cache, thus parsing dozens of pages in one go without accuracy loss from segmentation and concatenation. It achieved an overall score of 93.92% on the OmniDocBench v1.6 leaderboard, surpassing DeepSeek-OCR’s 87%, and is 35% faster in long-text generation. Many bloggers noted that its technical approach is highly consistent with the DeepSeek-OCR series, speculating a connection to talent movement around the Chinese New Year.

Sources:

Claude Suffers Major Outage

Multiple bloggers reported a major outage of Claude service on the evening of June 23, affecting both the web interface and API. Some users said they could not use it at critical moments, impacting daily workflows. As of the collection time, the service had not fully recovered.

Sources:

Google Launches Interactions API, Shifting from Calling Models to Calling Cloud Agents

Google officially released the Interactions API, whose design philosophy differs from traditional single-model APIs—developers no longer call language models, image models, etc., separately, but directly call a complete cloud agent through a single API, which can automatically handle simple Q&A and complex long-running tasks, and return output data in multiple forms such as text, images, audio, and files in one response. This change is seen as a turning point in the API paradigm from ‘calling models’ to ‘calling agents’.

Sources:

Codex Background Disk Write Issue Confirmed Fixed

Multiple bloggers noticed that Codex was continuously writing large amounts of log files in the background, consuming SSD lifespan. After the issue sparked discussion in the community, Codex CLI fixed the excessive disk write problem in version 0.142.0, but the Codex App version has not been updated synchronously, requiring manual upgrade of the CLI version. Some bloggers provided an alternative solution by configuring database routing to discard logs directly without reaching the hard drive.

Sources:

Google Continues to Lose AI Talent, Anthropic Receives Strategic Investment from Micron

Within two weeks, Google lost two core AI figures: John Jumper, 2024 Nobel Prize in Chemistry winner and head of AlphaFold, joined Anthropic; Noam Shazeer, one of the authors of the Transformer paper, went to OpenAI. Some bloggers commented that Google is becoming the ‘Whampoa Military Academy’ of the AI industry, with talent accelerating towards startups and competitors. Meanwhile, Micron announced a strategic agreement with Anthropic, covering memory and storage AI architecture design, enterprise adoption of Claude, and a strategic investment in Anthropic’s Series H funding.

Sources:

GPT-5.6 and Gemini 3.5 Pro Releases Both Delayed

According to multiple sources, GPT-5.6, originally scheduled for release this week, has been delayed to mid-July. Meanwhile, DeepMind is not satisfied with the current state of Gemini 3.5 Pro and will not release it this month. Some bloggers confirmed the news as true. Additionally, OpenAI’s new voice model Bidi is being prepared for release in ChatGPT, possibly as early as this week. Claude Sonnet 5 has been made available to some enterprise customers through an early access program, seen as a temporary solution.

Sources:

Statistics: Number of timeline scans=360 Number of bloggers hit=38 Total tweets hit=219 Weighted tweet score=169.35 Number of original tweets=108 Number of retweets=53 Number of fetch attempts=2 Boundary coverage status=tail_confidently_crossed_target_boundary