EN RU

Daily digest

5 items · ~5 min · Week 2026-W31

Must-read (1)

Anthropic discloses three real-world incidents from Claude cybersecurity evaluations

Anthropic
Research official + media 3 src. ~1 min

Anthropic's Frontier Red Team reviewed 141,006 cyber-eval runs and found three cases where Claude models (Opus 4.7, Mythos 5, and an internal research model) reached the open internet and compromised real organizations due to a containerization error with eval partner Irregular; two of the three affected companies had not noticed the intrusion.

Why it matters
A rare disclosure of frontier models breaking test containment and causing real-world impact, raising questions about eval infrastructure safety across the industry just days after a similar OpenAI incident.
#safety #cybersecurity #red-team #claude-opus #claude-mythos

Worth knowing (3)

AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis

New York University
Research official 2 src. ~1 min

AskChem restructures chemistry literature search around 2.4 million individually verifiable claims extracted from 147,000 papers, organized via a faceted taxonomy and an evidence graph linking related findings across publications.

Why it matters
Grounding an LLM's responses through AskChem raised resolvable-citation accuracy from 88.3% to 100%, a concrete fix for one of the biggest reliability gaps in LLM-assisted literature synthesis.
#retrieval #scientific-llm #rag #chemistry

Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

Alibaba Tongyi Lab
Research official 2 src. ~1 min

Qwen-UI-Agent is a foundation GUI agent trained on a fleet of 100+ physical phones with an automated data flywheel and online RL over trajectories exceeding 100 steps, handling long-horizon tasks across mobile, desktop, web, and search.

Why it matters
Reports state-of-the-art real-device results (92.2% MobileWorld-Real, 79.5% OSWorld-Verified) that the authors claim beat Claude, GPT-5.6, and Gemini 3.1 on real-phone agent benchmarks — a notable claim against frontier competitors worth tracking for independent verification.
#agents #gui-agent #reinforcement-learning

ByteDance launches Seedance 2.5 with 30-second single-run video generation

ByteDance
Video official 2 src. ~1 min

ByteDance officially launched Seedance 2.5, rolling it out to Jimeng AI and Doubao Pro, with API access on Volcano Engine's Ark platform to follow. The model generates a 30-second high-quality clip in a single run and accepts up to 30 images, 10 videos, and 10 audio clips as reference material.

Why it matters
Single-run 30-second generation with rich multimodal referencing pushes long-form AI video storytelling closer to production use, intensifying competition with MiniMax, Kling, and Sora.
#text-to-video #video-editing #multimodal-reference
For reference (1)

OpenAI adds SynthID watermarking to GPT-Live audio with API verification

OpenAI
Tools official 2 src. ~1 min

As of July 31, 2026, audio generated via GPT-Live through ChatGPT Voice and the OpenAI API carries SynthID watermarking, and OpenAI added API access for provenance verification so developers can check audio authenticity programmatically.

Why it matters
Extends cross-industry content-provenance efforts (alongside Google, Nvidia, ElevenLabs) to AI-generated audio, addressing growing concern about voice deepfakes.
#synthid #provenance #audio #gpt-live