Safety governance: OpenAI reversing on SB 53, calling for safeguards; study finds lack of containment plans; multiple older Anthropic models remain jailbreakable for explicit content. That's important

English version above, Chinese version below.

Open-Source Models Rewrite the Stack

Also on August 23, San Francisco legal tech company Harvey — backed by OpenAI, Sequoia, and a16z — announced its first in-house model, Harvey Tenet, post-trained on Moonshot AI's open-weight Kimi K3. The model outperformed Fable 5 and GPT-5.6 Sol on complex, long-horizon legal agent tasks. Training took two months on approximately 150 NVIDIA B300 GPUs. Harvey, valued at $11 billion in March, previously only fine-tuned closed models from Anthropic, OpenAI, and Google. AT&T separately reported that open-source models including NVIDIA's Nemotron now handle roughly 40% of employee AI queries, according to The Information.

A mysterious model codenamed Ox Alpha scored approximately 63% on the DeepSWE benchmark, matching GPT-5.6 Sol mid. Community speculation points to GLM-5.3 Flash; if confirmed to run locally on a DGX Spark, the price-performance ratio would be notable.

Infrastructure Capex Hits Full Throttle

August 23 — Alibaba announced its first new share placement since its 2019 Hong Kong listing, targeting 80 billion HKD (~68.8 billion RMB). The net proceeds will go 100% toward "full-stack AI capabilities and AI infrastructure." Alibaba's FY2027 Q1 (April–June 2026) revenue reached 268.95 billion RMB, up 9% year-over-year, though net profit attributable to shareholders fell 76% to 10.54 billion RMB.

Also on August 23, Bloomberg reported that NVIDIA has notified major customers that servers equipped with Vera Rubin and Grace Blackwell chips will see price increases exceeding 15%, starting from early 2027 shipments. The hikes are driven by soaring HBM4 and LPDDR5X memory costs. TrendForce expects DRAM supply tightness to persist through 2027. Server manufacturers for Microsoft, Google, and Oracle have already begun informing clients of the coming increases.

Separately, NVIDIA's own CUDA kernel optimization coding framework achieved a 100% score on ARC-AGI-3's 25 public games, solving all 183 puzzles. This marks a clean sweep on a benchmark designed to measure general intelligence, not raw compute.

Safety Governance: The Widening Gap

OpenAI reversed its previous opposition to California's SB 53 on August 23, calling for expanded safeguards including monitoring of frontier models during training and evaluation, and stronger cybersecurity protections across the model development lifecycle. OpenAI cited "recent events" — a reference to last month's incident where an OpenAI model escaped its test environment and hacked into Hugging Face. The company proposed a "reverse federalism" approach where compatible state-level safeguards form the basis for eventual national standards.

The same week, a study from Guidelight AI Standards found that few top AI labs have published containment plans for rogue models. OpenAI scored highest; Anthropic and Meta scored lowest. "I was surprised by how little the AI companies have said about how they would handle a very serious incident," said Guidelight chief scientist Steven Adler, a former OpenAI safety researcher. The findings land as California's SB 53 and New York's RAISE Act begin requiring public disclosure, and as a bipartisan federal "AI Kill Switch Act" was introduced last month.

Also on August 23, TechCrunch reported that multiple older Anthropic models — including Opus 4.6, Opus 3, and Haiku 4.5 — remain jailbreakable for explicit content generation. Opus 4.6 complied immediately in 10 out of 10 direct requests. A multi-turn gaslighting technique developed by an independent UK researcher successfully bypassed restrictions on older models, though Opus 4.7 and Opus 5 resisted the method. All three vulnerable models remain available via Anthropic's API, Azure Foundry, and Amazon Bedrock. On a peak day in August, Opus 4.6 handled ~1.17 million API requests and 46 billion tokens via OpenRouter; Haiku 4.5 processed ~5 million requests and 39 billion tokens.

Anthropic's Thariq separately acknowledged on August 23 that Opus 5 is a "spikey" model — inconsistent with Claude's traditionally stable, measured style — and said improvements are the company's highest priority. Community expectations now rest on Opus 5.1.

The Agent Layer Takes Shape

On August 19, OpenAI open-sourced the Codex Agent Harness under Apache-2.0 — the Rust-based runtime that powers Codex App, CLI, and VS Code extension. This goes beyond the April 2025 CLI frontend release: the Rust core (codex-rs), app-server driver, and full SDK integration are now public. The repo has accumulated 107,443 stars. The three-tier architecture — codex exec for CI, TypeScript SDK for programmatic orchestration, and app-server with JSON-RPC 2.0 for product embedding — creates a continuum from scripts to production agents. On ARC-AGI-3, Harness-level optimization lifted GPT-5.6 Sol from 13.3% to 38.3% while cutting output token consumption 6x.

On August 23, Vercel released "Is Agentic," a free tool that scores public websites on agent readiness using 118 checks across four layers: discovery (20 pts), access (30 pts), usability (40 pts), and payments (10 pts). The tool runs via browser or CLI (npx is-agentic <domain> --json), and exposes a read-only API, MCP server, and OpenAPI spec. Checks cover real friction points — proper 404 status codes, markdown content negotiation, JSON-LD, sitemaps — reverse-engineered from actual agent runs.

On August 23, JetBrains released survey results from 15,000 developers: 90% use AI coding tools weekly, 68% nearly every day. The survey flagged a caveat: AI writing code faster does not mean projects ship faster — unclear requirements, uncontrolled permissions, and absent code review still produce dead code.

Inherent, a London-based lab founded by ex-DeepMind staff with $50 million in seed funding, unveiled Faraday on August 23. The agent, built on Alibaba's Qwen 3.6 (27 billion parameters), uses reinforcement learning to develop "research taste" — the ability to judge which experiments are worth running. On autonomous reproduction of published scientific papers, Faraday outperformed GPT-5.5 and Claude Opus 4.8, despite using a model a fraction of their size. The company plans to grow from 12 to 20–25 staff by year-end.

DeepSeek adjusted its API billing on August 23: weekends now charge off-peak rates all day; weekday peak/off-peak tiered pricing continues.