[AINews] not much happened today

Anthropic rolled out Projects in Claude Code, allowing a single conversation to spawn parallel cloud sessions, pass context between threads, and keep running after the user leaves. Follow-up posts stated threads currently run in the cloud, with local workflows coming. Anthropic staff described it internally as a higher-level coordinator abstraction with evolving long-lived memory and aggregated status updates via a single controlling Claude (Cat Wu, MikeyK).
OpenAI launched Astra for Law with 26 partner-built plugins and 47 community plugins, initially rolled out through Trusted Access in ChatGPT and Codex, with API access to come later. Vals said reported runs show Astra for Law beating generic GPT-6 Astra plus web search on its legal benchmark at every price point.
Google updated Gemini managed agents with a new Antigravity-based harness plus a Credentials API that keeps secrets out of model context via placeholders and trusted-domain egress proxying, and a Files API for artifact movement and persistent sandboxes. The release claims up to 30% lower costs and 22% higher cache hits.
TypeSafe's Jev drew discussion as a fast, cheap constrained-output primitive used for LLM-as-judge, harness routing, subagent creation, and structured outputs. LangChain noted Jev is useful because it is not meant for free-form generation. Cloudflare exposed it via AI Gateway. Theo argued that using Jev for aggressive line-by-line history compaction misunderstands agent memory, reasoning traces, and cache economics, adding that editing history can cost more than leaving it because it invalidates cached prefixes.
A security incident drew attention: multiple posts summarized WSJ reporting and the researchers' own writeup that three researchers used Claude Opus 5 to chain an image-upload bug, ChatGPT/Codex account takeover, and access to OpenAI-connected services, proving it with a PR in OpenAI's internal monorepo, reportedly in under 72 hours and for under a few thousand dollars in tokens (Yuchen Jin, WSJ).
Anthropic published three measurements for tracking AI development: how much AI R&D is done by AI, how well agents are overseen, and how compute is allocated. Secondary discussion cited a Claude-led share of model R&D tasks rising from 1% to 26% in about six months, over 90% of model R&D work involving Claude collaboration or leadership, and about 30,000 internal agents active. Anthropic also said Claude optimized inference for 30+ open-source biology models, averaging 4x speedups, with code open-sourced.
Google's DeepMind published Stellar Colosseum, a model-agnostic many-agent harness for mathematics and TCS. Claimed results include a Codeforces 4263 and 71.0% on TCS-Bench. NVIDIA-associated work on Agora uses Git commits as shared memory for 13 workers over 12 days. Epoch launched Benchmark Reviews with 15 audits labeled Verified, Flawed, or insufficiently documented.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
a quiet day