Morning Brief: Sunday, June 28

Sixty-six feeds. Two weeks. 4,800 items reduced to what follows. (what we track, how we crawl, subscribe)

Sunday. No arXiv (weekend). The week resolves around one frame: both frontier labs released their next-generation models under US government allocation. OpenAI shipped GPT-5.6 Sol / Terra / Luna to trusted partners Friday and put up a public Deployment Safety Hub for GPT-5.6 Preview overnight. The administration separately authorized Anthropic's Mythos for ~100 US companies and agencies. Two labs, one allocation pattern. Restricted release is no longer the OpenAI story — it is the release model.

The international response started immediately. TechCrunch reports Asian AI startups launching Mythos-like models as the export ban drags on. The thread Schneier opened two weeks ago on whether state allocation produces parallel infrastructure has its first data point.

Underneath the geopolitics: developer token economics. Slashdot ran a piece this morning arguing developer AI token costs could exceed their salaries within two years. Latent Space reported Friday that OpenAI's median internal Codex output tokens grew 56x in Research, 32x in Customer Support, 27x in Engineering, and 13x in Legal since November 2025. The same usage curve that makes harness eval matter at GitHub/Vercel scale is what makes the salary-versus-tokens math start to bite at line-engineer scale.

Quieter but load-bearing: The New Stack on context debt as the disease behind vibe slop; on Greptile/Cursor/Devin converging on agents running their own code; "Post-Mythos Cybersecurity: Keep calm and carry on" picked up on HN; a Slashdot piece on how a harmless image can jailbreak vision-language models.

Top (5-7 min)

GPT-5.6 Preview System Card — OpenAI Deployment Safety Hub: Pinboard, 2026-06-28. The first-party artifact for the restricted-release pattern. OpenAI now publishes a public safety hub keyed to the preview release that the US government must approve for downstream access. Closes the loop on Friday's Sol/Terra/Luna announcement and yesterday's METR predeployment evaluation.
US Government Allows Anthropic Limited Release of 'Mythos' AI Model: Slashdot, 2026-06-27. Plus TechCrunch: Trump Admin releases Anthropic Mythos to be used by more than 100 US companies, agencies. Restricted release is now bilateral. Both frontier labs ship through the same USG allocation channel. The Fable suspension story that started June 9 ends here.
Asian AI startups launch Mythos-like models: TechCrunch, 2026-06-27. The international response begins immediately. Export controls produce parallel infrastructure; the only question was how fast.
Developer AI Token Costs Could Exceed Their Salaries in Two Years: Slashdot, 2026-06-28. Pairs with Latent Space: OpenAI internal Codex output tokens grew 56x in Research since November 2025. The cost curve that the platform vendors are building harnesses to control is now visible to line engineers.
Vibe slop is the symptom. Context debt is the disease.: The New Stack, 2026-06-27. Names the cost beneath generated code: context the human reviewer doesn't have. Sets up TNS: Greptile, Cursor, and Devin agree agents should run their code — what agents run the code against becomes the load-bearing question.
Post-Mythos Cybersecurity: Keep calm and carry on: HN, 2026-06-27. Cyber-defense practitioner reads the Mythos restriction as a non-event for daily defensive work. Useful counter to the alignment-discourse framing.

Themes this week

Restricted release becomes bilateral: GPT-5.6 Sol/Terra/Luna restricted to trusted partners (Fri), METR predeployment evaluation of Sol (Fri), TNS: US government just told OpenAI who's allowed to use GPT 5.6 (Fri), USG authorizes Anthropic Mythos for trusted orgs (Sat), TechCrunch: ~100 companies and agencies on the Mythos allow-list (Sat), OpenAI Deployment Safety Hub for GPT-5.6 Preview (today), Asian AI startups launch Mythos-like models (Sat). Both labs, one allocation channel, parallel international response.
Token economics meet salary math: Developer token costs may exceed salaries in two years (today), OpenAI's own Codex usage: 56x Research, 32x Customer Support, 27x Engineering, 13x Legal since Nov 2025 (Fri), TNS: How we cut AI costs by 80% (Thu), GitHub Copilot harness eval methodology (Thu). The vendor harness story has a line-engineer cost story under it.
Meta-Harness Summer, vendor-side codification: GitHub's Copilot harness evaluation methodology, Latent Space: Meta-Harness Summer, TNS: session as the new unit of compute, Vercel AI SDK 7 + harness adapters, TNS: Greptile/Cursor/Devin on agents running their code (yesterday). Five vendors converge on the same abstraction.
Akrites and maintainer-coordination response: "Akrites" vulnerability-mitigation project launches (LWN), TNS on the 19-org coalition, open letter + LF sponsorship.

Scan (15 min)

Agents, harnesses, model releases
- GPT-5.6 Preview System Card — Deployment Safety Hub, OpenAI, 06-28
- GPT-5.6 Sol/Terra/Luna restricted release, Latent Space, 06-27
- USG allows Anthropic limited Mythos release, Slashdot, 06-27
- Trump Admin releases Mythos to 100+ orgs, TechCrunch, 06-26
- Asian AI startups launch Mythos-like models, TechCrunch, 06-27
- METR predeployment eval of GPT-5.6 Sol, METR, 06-26
- GitHub Copilot harness eval methodology, GitHub Blog, 06-26
- Meta-Harness Summer, Latent Space, 06-25
- OpenAI internal Codex output tokens: 56x growth since Nov 2025, Latent Space, 06-26
Cost, context, and slop
- Developer AI token costs may exceed salaries in two years, Slashdot, 06-28
- Vibe slop is the symptom. Context debt is the disease., The New Stack, 06-27
- Greptile, Cursor, Devin agree agents should run their code, The New Stack, 06-27
- Your engineering org needs an AI slop registry, The New Stack, 06-26
Corp engineering and runtime
- Saga rollbacks for Cloudflare Workflows, Cloudflare, 06-25
- AWS, Microsoft, Google: session as the new unit of compute, The New Stack, 06-26
- The AI agent identity problem nobody's talking about, The New Stack, 06-26
- How I hunt for vulnerabilities with AI, ClickHouse, 06-26
- Trace and debug eve agent sessions with Vercel Observability, Vercel, 06-26
- Run a vLLM Server on HF Jobs in One Command, Hugging Face, 06-26
Eval, safety, governance
- Deployment Awareness Matters More Than Evaluation Awareness, Alignment Forum, 06-26
- The Case for Model Forensics, Alignment Forum, 06-26
- What happened after 2,000 people tried to hack my AI assistant, Simon Willison, 06-26
- Incident Report: CVE-2026-LGTM, Simon Willison, 06-26
- Post-Mythos Cybersecurity: Keep calm and carry on, HN, 06-27
- How a seemingly harmless image can jailbreak vision-language AI models, Slashdot, 06-27
Akrites, supply chain, CVE coordination
- "Akrites" vulnerability-mitigation project launches, LWN, 06-26
- After Fable 5 ban, Anthropic + 19 organizations launch OSS security body, The New Stack, 06-26
Surveillance and critique
- Meta Is Testing Facial Recognition for Police and Military, Schneier, 06-26
- One Million Passports Leaked Online, Schneier, 06-26
- Zuckerberg's increasingly bizarre war on whistleblowers, Pluralistic, 06-27
- Lawmakers Must Act Now to Prevent Armed Police Drones, EFF, 06-26
Systems, BSD, kernel
- Three stable kernel updates, LWN, 06-27
- Lots of stories about systemd v261, LWN, 06-26
- What's coming in Git 2.55, LWN, 06-26
- Reports from OSPM 2026, day three, LWN, 06-26
Aviation, defense, space
- First U.S. deployment of a collaborative combat aircraft, The Air Current, 06-26
- Bjorn's Corner: Aircraft Structures Part 7. Fiberglass, Leeham, 06-26
- Scroll burned in 79 AD volcanic eruption deciphered using AI, Slashdot, 06-27
- SpaceX plans 'Starpipe' natural gas pipeline for Starship, Slashdot, 06-27
Clojure, OCaml, languages
- Inside chachaml: Core Capabilities for AI-Native Workflows in Clojure, Planet Clojure, 06-26
- OCaml 5.5 Summary, OCaml.org, 06-23
- Neocaml 0.9: a better REPL, Dune/Opam completion, OCaml.org, 06-24
Research
- Accelerating Gemini Nano on Pixel with frozen Multi-Token Prediction, Google Research, 06-26
- After 80 Years, Mathematicians Give Famed 'Erdős Method' an Upgrade, Quanta, 06-26
wal.sh first-party
- Latent Configuration: When Deploy != Activate, wal.sh, 06-27
- Code Search and Code Graph MCP Servers on FreeBSD: A 2026 Survey, wal.sh, 06-27
- CIDER 1.22: Sao Miguel, wal.sh, 06-27
- The Agent Context Thread: An Index, wal.sh, 06-26

Tail

Scientists think they've uncovered the 15-million-year-old origin of laughter, 404 Media, 06-27
Beg, borrow and steal, Kenneth Payne, 06-26
Heat waves mess with your brain. Scientists are trying to figure out why., MIT Tech Review, 06-26
Hate "The Algorithm?" RSS Is One of the Tools You've Been Looking For, EFF, 06-26

Feed silences (diagnostic)

arxiv-cs-ai: weekend pause, returns Monday.
Anthropic first-party RSS: still 404; covered via anthropic-generated + Latent Space + TNS + Slashdot/HN aggregation.
bitsavers (6 feeds): connected, 0 items (sparse archive).
James Bornholt, Netflix Tech Blog: persistent DNS/TLS errors.
claude-auth: brief generation auth failed every morning June 21–28. Today's brief regenerated manually.

Build provenance

build: 2026-06-28 | crawler-sha: 5fe7ab8 (Walsh-Research/1.2, compliance v1.3) | feeds: 66 active (78 configured, incl. 12 corp-eng, 6 bitsavers, 5 generated) | items-considered: 4800 (14d, incl. 2939 arXiv) | warehouse: 19707 items | published: 48 | note: GPT-5.6 Preview System Card published on OpenAI Deployment Safety Hub; USG authorizes Anthropic Mythos for ~100 trusted orgs; Asian AI startups launch Mythos-like models; restricted release now bilateral across OpenAI + Anthropic; developer token cost / Codex usage curve becomes visible