Daily AI Digest

One transmission fault dropped 3.1 GW of data center load at once, and PJM needed 11 minutes to stabilize

After a transmission line fault in Northern Virginia, nearby data centers switched to backup power within about 30 seconds and removed roughly 3.1 GW of demand from the grid almost simultaneously. PJM data showed surplus power peaking at 3.49 GW, with the system taking about 11 minutes to settle. The event did not cause a major outage, but it exposed the synchronization risk created when many neighboring facilities use the same protection thresholds. Staggered disconnect and reconnect behavior, or campus batteries that carry compute through brief disturbances, could absorb the shock instead of passing it directly to the public grid.

Verdict: Teams building or buying AI capacity should put grid ride-through requirements into their contracts, or compute growth will turn local faults into system-wide risk.

Thibault Sottiaux of the OpenAI Codex and ChatGPT team says ChatGPT Work is now available worldwide on every paid plan across mobile, web and desktop. He describes it as putting a jetpack on ChatGPT. For product teams, the signal is bigger than one more surface: general chat products are making persistent task execution a default experience.

Zara Zhang, an independent builder, says what she wants most from models now is not more intelligence but faster responses. A wait of one to five minutes is too short for deep work and long enough to push a user back into the feed, leaving agents with an interaction cost that can outweigh better benchmark scores.

Peter Steinberger, creator of OpenClaw, says his automated review skill ran for 66 rounds on a complex refactor, a personal record. The run shows how repeated verification can push agents into deeper engineering work, but it also argues for convergence criteria, budgets and explicit exit signals so repeated edits are not mistaken for progress.

Anthropic cut more than 80% of Claude Code's system prompt for Claude 5: fewer rules, better interfaces

Anthropic says it removed more than 80% of Claude Code's system prompt for Claude Opus 5 and Fable 5 with no measurable loss on its coding evaluations. Its updated practice moves from stacking rules to letting the model use judgment, from tool-call examples to clearer interfaces, and from putting everything up front to progressive disclosure through skills, deferred tool loading and automatic memory. The company has also put these checks into /doctor to help shrink CLAUDE.md and skill files. The Hacker News thread reached 139 points and 88 comments in the collection window, with debate over whether stronger models reduce prompt engineering or merely move the complexity into tools and evaluations.

Kimi K3 trails frontier closed models on cyber tasks, but completed a full corporate attack chain once in ten runs

A preliminary UK AISI and U.S. CAISI evaluation scored Kimi K3 at 32% on ExploitBench, above GLM-5.2 at 24%, but it reached arbitrary code execution in none of 41 vulnerability samples, compared with an average of 20 for leading closed models. On the 32-step simulated corporate network called The Last Ones, Kimi K3 reached step 17 on average, versus 28.5 for the leading U.S. models, yet it still completed the entire chain in one of ten attempts. Evaluators also found that its safeguards did not prevent assistance with exploit development. These are preliminary results from limited benchmarks, with system-level refusals disabled for some comparison models, so they are not a direct estimate of real-world attack rates.

Verdict: Security teams preparing to deploy open-weight agents should design isolation around the worst successful run, not the model's lower average capability.

Debian weighs the boundaries of AI-assisted contributions: licensing, disclosure and final accountability

Debian developers are discussing a general resolution on whether LLMs and generative AI can be used in project contributions. The permissive proposal requires contributors to check tool terms and third-party licenses, accept full responsibility for technical quality, security and maintainability, and disclose substantial tool-generated content with markers such as Generated-By or Assisted-By. Bulk automated changes would need prior discussion and human oversight. A competing proposal calls for a ban, while other comments seek narrower limits, so none of the drafts should yet be treated as Debian policy.