10 Jul 2026

Digest: GPT-5.6 general release, AI Futures' Plan A, dual-use 'off switch'

OpenAI pushes GPT-5.6 and a long-horizon workplace agent into general release, while the safety side of the field has a consequential day of its own: a superintelligence-governance blueprint from the AI 2027 team, a pretraining-time 'off switch' for dangerous knowledge from Anthropic, and $160M for independent alignment research.

1. GPT-5.6: Frontier intelligence that scales with your ambition

   OpenAI's GPT-5.6 family — Sol, Terra and Luna — is now rolling out to all users across ChatGPT, Codex and the API, ending the staged release that began in late June with previews and a government-approved customer tier. OpenAI reports state-of-the-art results on agentic benchmarks — 80.0 on the Artificial Analysis Coding Agent Index (a third-party measure of coding-agent performance), 2.8 points above Claude Fable 5 at under half the output tokens — though launch-day figures are vendor-selected and independent testing of the released models is still to come; it is the latest move in the agentic-coding race that has dominated frontier competition in recent months.

   OpenAI
2. AI 2040: Plan A Recommended

   The AI Futures Project — the team behind the AI 2027 scenario — published 'AI 2040: Plan A', a year-in-the-making blueprint for what they think should happen rather than what will: international coordination that deliberately delays superintelligence from roughly 2030 to 2040 to buy time for alignment. The authors call it 'the least bad plan we currently know of', and notably launched it alongside commissioned critiques, with safety researchers immediately split on whether it is workable or goes far enough.

   Daniel Kokotajlo via AI Futures Project
3. Jul 8, 2026 Alignment An off switch for dual-use knowledge in AI models Recommended

   Anthropic and AE Studio unveiled GRAM (gradient-routed auxiliary modules), a pretraining technique that routes dual-use knowledge — virology, cybersecurity and nuclear physics in their tests — into dedicated, removable compartments, so deleting a module strips the capability about as thoroughly as never training on the data, while trusted users could have it switched back on. Unlike post-hoc 'unlearning' (removing knowledge after training, which attackers can often reverse), the isolation reportedly survives adversarial fine-tuning — but the work is preliminary: demonstrated only up to 5B-parameter models, not applied to production systems, and cleanly separating entangled knowledge (general biology vs dangerous virology) remains an open problem.

   Anthropic Research
4. Announcing our $160M grant from Coefficient Giving

   Resolution — the alignment research organisation that launched last month as Sequent, founded on the argument that alignment research is 'not on track' — announced a $160M grant from Coefficient Giving, structured as a $108M base plus $52M conditional on hiring and compute needs, to put rigorous alignment research on a closer-to-even footing with the frontier labs. An unusually large sum for independent safety research, it suggests safety philanthropy beginning to scale toward lab-sized budgets.

   Geoffrey Irving, Resolution (fka Sequent) via Alignment Forum
5. ChatGPT is now a partner for your most ambitious work

   Alongside GPT-5.6, OpenAI launched ChatGPT Work, an agent powered by Codex and GPT-5.6 that takes actions across a user's apps and files and can stay with a project for hours, turning a stated goal into finished work. It moves long-horizon autonomous operation — until recently mostly an evaluation construct — into mainstream workplace deployment, substantially widening the population of unattended agent-hours.

   OpenAI
6. Australian minister details AI Safety Institute progress at 2026 AI Safety Forum

   At the AI Safety Forum in Sydney, Industry Minister Andrew Charlton said Australia's new AI Safety Institute is now testing frontier models, has hired a Safety Science Research Lead and staff from the UK AI Security Institute and Google DeepMind, and has signed information-sharing agreements with the UK and Canadian safety institutes. It extends the steady build-out of the international evaluation network that has been a recurring governance theme in recent weeks.

   Andrew Charlton via Department of Industry, Science and Resources
7. Because 8 ≈ e², Anthropic's researcher uplift is plausibly >2x

   In a short modelling note, METR's Thomas Kwa argues that Anthropic's report of contributors merging 8× as much code per day as in 2021–24 plausibly implies each researcher's effective output has more than doubled (a 'researcher uplift' above 2×). The note is candid about its limits — it is one researcher's opinion, colleagues at METR disagree, and code volume is a rough proxy for research — but it is a rare outside attempt to quantify how much frontier-lab R&D is already accelerated by the labs' own models, the recursive-improvement dynamic safety researchers watch most closely.

   METR