Morning Brief: Saturday, June 27

Seventy-eight feeds. Two weeks. 5,532 items reduced to what follows. (what we track, how we crawl, subscribe)

Saturday. The week's open question gets its answer overnight: OpenAI ships GPT-5.6 as a restricted-release family (Sol / Terra / Luna), trusted-partners only. METR publishes its predeployment evaluation of GPT-5.6 Sol the same morning, and The New Stack covers what amounts to the first explicit US-government gating of an OpenAI release. Tuesday's White-House stagger ask, yesterday's "must be punished" Alibaba escalation, and the Akrites launch resolve into one frame: model access is now a state-coordinated allocation, not a product launch.

Meta-Harness Summer gets its platform receipt. GitHub publishes the methodology behind its Copilot agentic harness evaluation across models and tasks — vendor-side codification of what Latent Space named two days ago. The harness, not the model, is the comparable unit.

The Akrites thread widens from yesterday's HN letter to LWN ("vulnerability-mitigation project launches") and The New Stack ("19 organizations launch open source security body"). The maintainer-versus-scanner asymmetry now has an industrial coalition.

Quieter but load-bearing: Cloudflare's saga-rollback writeup for Workflows (correctness primitive for distributed agent state), The New Stack on AWS/Microsoft/Google converging on "session as compute" but diverging on isolation, Schneier on Meta testing facial recognition for police and military, The Air Current on the first US deployment of a collaborative combat aircraft.

Top (5-7 min)

OpenAI GPT-5.6 Sol / Terra / Luna — restricted to trusted partners: Latent Space, 2026-06-27. Pairs with METR's predeployment evaluation of GPT-5.6 Sol published the same day and The New Stack: US government just told OpenAI who's allowed to use the next GPT 5.6 model. Three signals — capability release, third-party eval, state-actor allocation — landing in one window.
Evaluating the GitHub Copilot agentic harness across models and tasks: GitHub, 2026-06-25. The harness eval methodology published by a platform vendor. Codifies what Latent Space framed two days ago as Meta-Harness Summer — the harness, not the model, is the unit of comparison.
"Akrites" vulnerability-mitigation project launches: LWN, 2026-06-26. The thread widens. Plus The New Stack: After Fable 5 ban, Anthropic and 19 organizations launch open source security body. Yesterday's letter (We All Depend on Open Source. We Will Defend It Together) was the announcement; today the maintainer press picks it up.
How we built saga rollbacks for Cloudflare Workflows: Cloudflare, 2026-06-25. The correctness primitive beneath distributed agent state. Compensating actions for partially-completed agent workflows — what makes async retry safe to ship.
AWS, Microsoft, and Google agree the session is the new unit of compute. They disagree on how to isolate it.: The New Stack, 2026-06-26. Three vendors converge on the same primitive — agent session as runtime — and split on the security model. The infrastructure question of the next 12 months in one paragraph.
What happened after 2,000 people tried to hack my AI assistant: Simon Willison, 2026-06-26. Simon's writeup of Fernando Iglesias's hackmyclaw experiment. Real-world adversarial data released the same week Patronus AI raised $50M for AI agent stress-testing.

Themes this week

Restricted release becomes the model release pattern: GPT-5.6 Sol/Terra/Luna restricted to trusted partners, METR's predeployment evaluation, The New Stack: US government allow-list, Simon Willison quoting OpenAI on the release, Deployment Awareness Matters More Than Evaluation Awareness (Alignment Forum). The pipeline — third-party eval, government allow-list, partner-only rollout — is what model release looks like now.
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, yesterday's Vercel AI SDK 7 + harness adapters. Four vendors converge on the same abstraction.
Akrites and the maintainer-coordination response: "Akrites" vulnerability-mitigation project launches (LWN), The New Stack on the 19-org coalition, yesterday's open letter + Linux Foundation sponsorship. The CVE-flood-vs-maintainer-capacity asymmetry now has institutional governance.

Scan (15 min)

Agents, harnesses, model releases
- GPT-5.6 Sol/Terra/Luna restricted release, Latent Space, 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
- Claude Tag: Multiplayer, Proactive, Persistent Agents in Slack, Latent Space, 06-24
- Introducing Claude Tag, Anthropic, 06-23
- Claude Code v2.1.195, claude-code-releases, 06-26
- A no-nonsense explainer to Agentic AI, Tailscale, 06-26
Corp engineering and runtime
- How we built saga rollbacks for Cloudflare Workflows, Cloudflare, 06-25
- AWS, Microsoft, Google: session as the new unit of compute, The New Stack, 06-26
- Silk: a silky smooth fiber runtime for ClickHouse, ClickHouse, 06-25
- Why we rewrote WAL-G for Postgres backups in Rust: WAL-RUS, ClickHouse, 06-25
- 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
- METR predeployment eval of GPT-5.6 Sol, METR, 06-26
- Deployment Awareness Matters More Than Evaluation Awareness, Alignment Forum, 06-26
- The Case for Model Forensics, Alignment Forum, 06-26
- AI and Liability, Simon Willison, 06-25
- Incident Report: CVE-2026-LGTM, Simon Willison, 06-26
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
- We All Depend on Open Source. We Will Defend It Together, Akrites, 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
- Lawmakers Must Act Now to Prevent Armed Police Drones, EFF, 06-26
- Jailbreaking isn't theft, Pluralistic, 06-25
Systems, BSD, kernel
- Lots of stories about systemd v261, LWN, 06-26
- What's coming in Git 2.55, LWN, 06-26
- Podman 6.0 released, LWN, 06-25
- Reports from OSPM 2026, day three, LWN, 06-26
- BSD Now 669: Poudriere Speed Run, BSD Now, 06-25
Aviation, defense
- The surprising 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
- Embraer's efficiency improvement comes from car-industry practices, Leeham, 06-25
Clojure and Scheme
- Inside chachaml: Core Capabilities for AI-Native Workflows in Clojure, Planet Clojure, 06-26
Research
- Accelerating Gemini Nano models on Pixel with frozen Multi-Token Prediction, Google Research, 06-26
- Which tokens does a hybrid model predict better?, Hugging Face, 06-25
- Optimizing cloud economics with linear elastic caching, Google Research, 06-25
- After 80 Years, Mathematicians Give Famed 'Erdős Method' an Upgrade, Quanta, 06-26

Tail

Quoting Dean W. Ball, Simon Willison, 06-26
Quoting Timothy B. Lee, Simon Willison, 06-26
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
Behind the Blog: Salesforce Beach, 404 Media, 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: 2,939 items in the 14-day window, fully live.
bitsavers (6 feeds): all connected, 0 items this run (sparse archive).
James Bornholt, Netflix Tech Blog: errors persist (DNS / TLS), permanent.
Anthropic first-party RSS: still 404, now covered by anthropic-generated
- Latent Space + The New Stack.
claude-auth: brief generation auth failed every morning June 21–27. Today's brief regenerated manually.

Build provenance

build: 2026-06-27 | crawler-sha: c04c1d1 (Walsh-Research/1.2, compliance v1.3) | feeds: 78 (incl. 12 corp-eng, 6 bitsavers, 5 generated) | items-considered: 5532 (14d, incl. 2939 arXiv) | warehouse: 19584 items | published: 53 | note: GPT-5.6 Sol restricted release + METR predeploy eval same morning; GitHub Copilot harness eval methodology; Akrites coverage widens to LWN/TNS; Cloudflare saga rollbacks; AWS/MSFT/Google "session as compute"; Meta tests facial recognition for police/military