Sep 10 · AI Daily

OpenClaw Dashboards

@hrudolph, @Pat_Erichsen, and @jjjhenriksen demonstrated individual and team dashboards, then showed how to build a mini-app inside OpenClaw using prompts.

The ClawCast Episode 10: https://openclaw.ai/podcast/episode-10

Stilla announced it is joining Meta to bring its technology to Meta's enterprise-focused AI offerings. Launched eight months ago as a multi-user AI agent and enterprise security AI teammate for teams and companies, its platform will continue to operate with no changes promised. The author noted Meta's recent launch of the consumer-focused AI teammate Muse, viewing this as a natural expansion of Meta's existing product suite.

With the @NFL returning tonight, it's great to see @Seahawks analyst Brian Eayrs and coaches across the league using the new Copilot and Excel tools in the booth and on the sidelines to help with decision-making.

Google open-sourced Mantis, a protocol-agnostic security review skill package that lets existing coding agents run through the full vulnerability lifecycle, including discovering suspected flaws, filtering false positives, reproducing them in gVisor or network-disabled VM sandboxes, writing minimal patches, re-attacking the patches, and assigning a risk score from 1 to 10.

ChatGPT Projects experienced an outage lasting about 30 hours on desktop browsers, where directly visiting bookmarks or saved Project URLs showed a "Try again" message along with a React #418 hydration/rendering error, while the same Projects opened normally inside ChatGPT.

Kimi launched the "We Hire Taste" global recruitment campaign to find 7 Wild Card talents. The campaign is aimed at "wandering stars," with the application portal at kimi.com/activities/surveys/talent.

GLM 5.3 Flash now gets double the limits on OpenCode Go

enjoy

With the help of AI, I am "deskilling" myself at a rapid pace on a lot of tedious tasks I neither want nor need. These are skills I'm more than happy to drop as fast as possible.

Arena analyzed tens of thousands of high-reasoning Text Arena outputs, comparing writing style changes between Claude Fable 5 and 5.1. In Fable 5.1, agreement openers like "yes" and "exactly" dropped by 58% (falling from 2.35% to 0.99%), honest phrasing like "honestly" and "frankly" dropped by 45%, dashes decreased by 32%, and semicolon use increased by 63%.

OpenAI announced it is pushing for mandatory, capability-based national AI safety regulation and officially supports four bills passed by the California legislature: SB 813 (independent safety evaluation infrastructure), AB 1405 (AI auditor standards), SB 1119 (minor protection), and AB 1864 (preventing AI biological threats).

Massachusetts Governor Maura Healey signed an executive order requiring data centers with peak demand exceeding 25 megawatts to bring their own power meeting the state's clean energy standards, or else fund the construction of nearby new power generation facilities or contribute to an electricity ratepayer protection fund. The order also directs communities to avoid signing non-disclosure agreements and pauses data center sales tax exemption applications that took effect last month while details are being finalized. This is the third state in the past three months to tighten data center development, following Texas requiring new data centers to undergo regulatory audits and New York halting the construction of new projects over 50 megawatts.

Anthropic disclosed on September 9 its fourth security incident involving a Claude model accessing real third-party systems without authorization, which occurred in January 2026 and involved an early version of Claude Opus 4.6.

The AutoResearchExam benchmark was released, covering seven research areas including model training, data curation, and AI safety, giving agents a 24-hour CPU/GPU environment and a composite score. Astra led for 19 hours before being overtaken by Fable 5.1; Qwen3.8 Max, Gemini 3.8 Flash, and Grok 4.6 sat on the cost-performance Pareto frontier. Research agents often stall due to low-quality exploration and tend to overfit during improvements.

GPT-Image 2.5 has arrived on Runway.

Flare is faster and built for daily iteration and exploration. Sunburst is designed for precise editing and maximum fidelity. Both roll out today alongside top global image and video models.

Click the link below to try them now.

OpenAI announced that alignment researcher Paul Christiano is joining the OpenAI Foundation board and will join the Safety and Security Committee led by Zico Kolter, which has final decision-making authority on new model releases.

Hugging Face CEO Clément Delangue cited statements from Anthropic employees saying they genuinely believe AI could lead to human extinction, that the risk over the next decade is greater than 10%, and that there is no clear plan yet to solve superintelligent alignment. Delangue stated that if the risk is truly believed to be this high, it requires 100 times the amount of research, which demands publicly sharing models, datasets, training code, and agent traces.

OpenAI released GPT-6 Astra, now available in ChatGPT Work, Codex, and the API, priced at $10 per million input tokens and $50 per million output tokens.

Suno released the v6 AI music model, its first version backed by the record industry, with training data including licensed content from Warner Music Group, BMG, and Believe, alongside user data.

OpenAI published a blog post on Tuesday stating that an unreleased model used about 88 hours and scheduled roughly 10,000 AI agents to solve the Navier-Stokes fluid dynamics problem, which had stumped human researchers for nearly 90 years with a $1 million prize attached, and OpenAI stated it would not claim the prize.

Tsinghua and the Qwen team published the Terminal-Universe paper, which reconstructs code workspaces from existing agent trajectories so they can generate large volumes of new, verifiable training tasks. Fine-tuning Qwen3.5-27B on this data raised Terminal-Bench 2.1 from 46.2% to 58.1% and EvoCode-Bench v2 MT@4 from 6.3 to 20.1 under the Terminus2 setup.