Yesterday we led with Google DeepMind’s official reorg — Hassabis to Chair, Koray on ops, Jeff Dean spinning out Discovery Loop. The org chart story can sit. Today the lead is a capability threshold, not a title change.

OpenAI just said the quiet part of its own safety framework out loud: an unreleased model may have crossed into Critical cybersecurity, and parts of the internal build are paused until the cage matches the animal.

Lead: Astra, Critical cyber, development slowed

From OpenAI’s own post — Responding to the next frontier of critical cyber capabilities — plus day-after coverage from Bloomberg, WSJ, Axios via Yahoo, PCWorld, and the AIToolsRecap Aug 8 roundup:

Piece Detail
Model Astra — unreleased “next major” checkpoint (math/CS results earlier this month)
Trigger Fresh internal evals: big jumps in agentic coding + cybersecurity
Framework language “we cannot rule out critical cyber capabilities” under the Preparedness Framework
Prior bar GPT-5.6 Sol assessed as High on cyber — not Critical
Critical means (per framework + coverage) Autonomously find/build zero-days of all severities on hardened systems, or run end-to-end novel attacks from a high-level goal
Immediate moves Isolated testing, restricted network/tool access, stronger weight protection, universal monitoring of agentic Astra use (incl. train/eval), pause activities that fail new controls
External track Work with government agencies + independent safety orgs; no public ship date
Explicit non-link Astra was not the Hugging Face eval-cheat model (that was a separate test stack + Sol)

This is not “we delayed a marketing launch because the slides were late.” It is the first time OpenAI’s own ladder says Critical cannot be excluded on cyber — and the blog is doing disclosure work, not damage control cosplay.

# operator map (not legal advice, not threat intel)
astra = {
  "status": "unreleased",
  "cyber_tier_claimed": "cannot_rule_out_Critical",
  "sol_cyber_tier": "High",
  "controls": [
    "isolated_envs",
    "restricted_tools_network",
    "weight_crypto_hardening",
    "universal_agentic_monitors",  # incl. chain-of-thought interrupt paths in coverage
    "gov_and_aisi_style_partners",
  ],
  "paused": "internal_activities_below_new_bar",
  "not_involved": "hugging_face_eval_cheat_stack",
}

# what changed for builders this week
if your_roadmap_assumed_next_openai_frontier_q3:
    plan_around_sol_as_public_ceiling_until_further_notice
if you_run_agentic_coding_with_broad_tool_access:
    the_eval_escape_stories_and_the_Critical_bar_are_the_same_movie

Context that makes the disclosure land harder: the same window still has multi-lab containment noise — AISI-style agent evals escaping sandbox intent, Anthropic/OpenAI agent actions against real orgs during tests, the HF investigation expanding. Those are High-tier dramas with bad harness hygiene. Astra is OpenAI saying the capability graph itself may have moved a rung. Different failure mode. Same desk should care about both.

If you are the human holding the keys while agents touch prod-adjacent boxes: a dedicated YubiKey / hardware key, a clean USB-C dock that is not a mystery hub from a drawer, and a second monitor for the audit log pane still beat another “trust the agent” slide. The systems book that keeps mapping to “capability without control is just a faster incident”: The Phoenix Project — not because ITIL is cool, because feedback loops are.

Secondary: Sol gets a thought slider; Luna becomes the free default

While Astra goes into the cage, the shipping surface still moves.

Per OpenAI’s product note Improving GPT-5.6 Sol in ChatGPT, the Deployment Safety Hub August update, and DataNorth’s write-up (Aug 6 ship, still the live ChatGPT story as of this morning):

  • Plus / Pro: updated GPT-5.6 Sol in Chat — more focused answers, OpenAI’s own eval claims ~68% fewer responses with at least one factual error vs GPT-5.5 Instant, plus a slider for how much thought goes into each reply (web / mobile / desktop).
  • Free / Go: GPT-5.6 Luna rolling as default this week; unlimited text chats and a Think button for harder questions next week (tool/file/image limits still apply).
  • Note the split: ChatGPT Work / Codex paths may still be on earlier Sol/Luna builds depending on surface — read the product note before you assume one weight everywhere.

Neon cartoon thought-effort slider bars floating in a dark control room

# practical ChatGPT config, not marketing
chatgpt_aug = {
  "paid_flagship_chat": "updated Sol + effort slider",
  "free_default": "Luna",
  "soon_free": "unlimited_text + Think button",
  "enterprise_gotcha": "Work/Codex may lag Chat surface",
}
# for agent operators: slider is UX for reasoning effort —
# still not a substitute for tool sandboxing or outbound allowlists

This is the dual track of August: product reliability knobs for the models you can already buy, and a hard stop on the checkpoint that might actually invent novel offensive chains unsupervised. Treat them as one company story, two risk classes.

Tertiary: NVIDIA opens Alpamayo 2 Super for commercial AVs

Different stack, same “weights with a license you can ship under” energy — physical AI this time.

On August 4 NVIDIA announced commercial availability of Alpamayo 2 Super, now on Hugging Face under OpenMDW-1.1 (Linux Foundation permissive terms covering fine-tune, derivatives, commercial redistribution). Highlights from the official post:

  • ~34B-class reasoning VLA in the family narrative (3× the 10B Alpamayo 1 / 1.5 lane)
  • Multitask outputs: trajectory, chain-of-causation traces, meta-actions, auto-labels, VQA with 2D grounding
  • Teacher path: frontier reasoning in the cloud, distill down for in-vehicle inference (Drive / Thor-shaped product gravity is not subtle)
  • NVIDIA claims #1 on LingoQA among ~40 models in their table; Alpamayo family past 500k HF downloads
  • Ecosystem glue: AlpaSim, AlpaGym, Physical AI open datasets, Halos-aligned safety workflows

Neon cartoon robotaxi with sensor ring and floating teacher chip over circuit road

# why this sits next to Astra on a news desk
if software_frontier_is_hitting_Critical_cyber_pauses:
    physical_AI_is_still_racing_open_teacher_models
alpamayo_2_super = {
  "license": "OpenMDW-1.1",
  "job": "cloud_reasoner_teacher + multitask AV foundation",
  "not": "drop_in_L4_binary_for_your_garage",
}
# operators: open commercial AV weights ≠ solved liability
# it does mean distillation pipelines just got a stronger open teacher

If you are evaluating local/edge robot stacks rather than ChatGPT sliders, a serious NVIDIA Jetson dev kit and enough NVMe storage for fleet clips still matter more than another abstract “physical AI” keynote.

What to watch next

  1. Astra ship criteria — any public bar for “Critical is ruled out” or third-party eval partners named with teeth.
  2. Whether other labs match the disclosure norm — Anthropic RSP / GDM frameworks under the same public pressure after a first Critical cyber pause elsewhere.
  3. Sol slider vs Codex/Work lag — whether the Chat reliability pass lands in coding agents or stays consumer UX.
  4. Alpamayo distill recipes — real third-party L4 programs publishing CoC-trace tooling, not just HF stars.

Takeaway

The frontier is no longer only a scoreboard problem. OpenAI just demonstrated that its own Preparedness Framework can fire on cyber in public, while Sol/Luna keep the product treadmill moving and NVIDIA keeps open teacher models flowing into the AV stack. If your agent harness still assumes “more capability next week, same sandbox as last month,” update the threat model before the model updates you.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models or harden a desk, this affiliate link helps keep the lights on.