Morning brief

Saturday, August 15

1 people · 0 companies · 4 topics

The signal

  1. Anthropic August 2026 Risk Report: safety benchmark saturated as R&D acceleration starts What happened: darioamodei / Anthropic published its second company-wide Risk Report on Aug 14; it upgraded misalignment risk from “very low” to “low,” revealed its CoBench internal safety benchmark has saturated and can no longer register incremental capability gains at the moment the company says it is seeing early signs of the R&D acceleration that threshold was built to catch, and disclosed an unreleased internal model (“Model 2”) measurably more capable than the externally-deployed Mythos 5. Why it matters: For an engineer building on Anthropic’s APIs, this is the first confirmation that the company’s own safety instrumentation is now blind at the frontier it is actively crossing — and that a more capable unreleased model already exists. Confidence: Confirmed Sources: Anthropic (primary) · Unite.AI (independent) · Benzinga (independent)

  2. Z.ai holds GLM-5.3 open weights after post-training produced full exploit chains What happened: Z.ai launched GLM-5.3 on Aug 14 with API access but is withholding open weights for roughly two weeks; post-training at scale caused cybersecurity capability to grow faster than Z.ai expected, with the model forming coherent multi-stage attack plans rather than spotting isolated bugs, and reaching 84.5% on CyberGym. Why it matters: First major Chinese frontier lab to voluntarily hold open weights because a capability emerged faster than safety hardening could track — a precedent that the dual-use crunch is arriving for open-source too, not just gated APIs. Confidence: Confirmed Sources: SiliconANGLE (independent) · Axios (independent) · TechTimes (independent)

Since last brief

  • model-releases Qwen3.8-27B shipped Aug 14 15:00 UTC under Apache 2.0, closing the companion-model gap the prior brief flagged as absent from the Qwen 3.8 Max release officechai.com

Models & research

Qwen3.8-27B

Claim: Alibaba released a 27.78B dense multimodal model (text + image + video) under Apache 2.0 with a 262K native context window. Evidence: Weights live on HuggingFace and ModelScope. Alibaba reports: SWE-Bench Pro 61.7%, Terminal-Bench 2.1 73.0 (vs Meta Muse Glimmer 30B at 51.7), LiveCodeBench v6 90.3%, GPQA Diamond 89.2%. All Alibaba-run benchmarks. Caveat: No independent reproduction yet. CoWorkBench comparison uses an older Opus 4.6 Max baseline. Model is bigger than Muse Glimmer’s 30B by a hair, which weakens the head-to-head framing. Worth your time: Read if you run local coding agents; Apache 2.0 and 24GB-VRAM fit make it a practical evaluation target.

Anthropic Model 2 (shelved)

Claim: The August 2026 Risk Report discloses an internal model Anthropic calls “Model 2” that is “somewhat more capable” than Mythos 5 across evaluated tasks; no external release is planned. Evidence: Self-reported in Anthropic’s own Risk Report. Model is not accessible outside Anthropic; no independent benchmarks exist. Caveat: “Somewhat more capable” is undefined. This is a voluntary disclosure with no quantitative detail attached — treat as a signal on direction, not a benchmark. Worth your time: Read (the disclosure paragraph, not the model). It updates your prior on the gap between Anthropic’s frontier and what the API exposes.

Needle 2 (Cactus Compute)

Claim: 45M-parameter, CQ2-bit quantized tool-calling model that ships as a 14MB binary and runs a full session in 28MB RAM on phones, wearables, and microcontrollers. Evidence: 500 tok/s on Raspberry Pi 5; 400-1500 tok/s on Meta Quest 3S and Apple Vision Pro. Cactus reports it trades wins with models 5-70x larger at full f16 precision on tool-calling benchmarks. Released ~Aug 13. Caveat: All benchmarks are Cactus’s own. CQ2-bit baked into training limits fine-tuning paths. Worth your time: Skim if you are building on-device or edge agents; otherwise skip.

Tools worth touching

Soup v3

  • Does: Fine-tunes and post-trains LLMs from a single YAML config using layer streaming, which pages decoder layers to VRAM one at a time, letting an 8B model train at 119.6 tok/s in 3.32 GB peak VRAM on a 4GB laptop GPU
  • Why now: v3 released Aug 13; supports SFT, DPO, GRPO, KTO plus eval, gating, and export in one CLI; no SSH, no distributed infra
  • Try if: You want to run local fine-tuning experiments on consumer hardware without a cloud bill
  • Skip if: Your target model exceeds 8B parameters or needs multi-GPU training
  • Links: GitHub · product page

Other

  • GLM-5.3 hit 28.3 on Terminal-Bench 3.0, up from 4.6 in GLM-5.2, on the same 743B base model via post-training only; if open weights hold to that when released (~Aug 28), it will be the strongest open-weight coding model available felloai.com
  • Anthropic’s August 2026 Risk Report also discloses that 133 million RLHF contractor exchanges (May 2025 to April 2026) ran without biological weapons classifiers active; risk rating for novel bio/chem uplift unchanged at “low, but higher than prior estimate” Anthropic (primary)

Watchlist

  • OpenAI ExploitGym / Kill Switch Act: still in House committee; no Senate companion bill (opened 2026-07-23)
  • Kimi K3 llama.cpp PR #26185: backend commits landed; mainline text-model merge status unclear (opened 2026-07-23)
  • White House AI security framework: confirmed will not be publicly released; watching for leaks via Congress (opened 2026-07-30)
  • EU AI Act enforcement: technical compliance dialogues ongoing; enforcement powers active since Aug 2; no formal restriction order or fine yet (opened 2026-07-26)
  • Anthropic eval breach / METR review: August 2026 Risk Report published Aug 14 with major findings (covered in Signal); METR independent review still pending (opened 2026-07-31)
  • OpenAI Astra: in 30-day government pre-release review; September is earliest plausible window (opened 2026-08-02)
  • Meta Muse Spark 1.1 breach: root cause confirmed; formal post-mortem and regulatory response still pending (opened 2026-08-06)
  • California Frontier AI Safety Act: passed Assembly Aug 10; awaiting governor signature by Sep 30 (opened 2026-08-11)
  • Demis Hassabis / Google DeepMind: Kavukcuoglu now SVP; Discovery Loop launched Aug 5; watching Gemini roadmap and talent pipeline impact (opened 2026-08-11)
  • Anthropic / Decart AI acquisition: $6B talks ongoing; no close or collapse reported (opened 2026-08-14)
  • Grok 4.7 (2.1T params): Grok 4.6 out Aug 12; Musk put 4.7 at “3-4 weeks” from Aug 12, tracking to early September; already a slip from the late-July window (opened 2026-08-14)