OpenAI split Daybreak into Blue and Red, and shipped GPT-5.6-Cyber — with hardware keys mandatory from today
OpenAI announced on August 31 that its Daybreak cyber-defense program now has two access tiers: Blue, built on GPT-5.6 Sol with the cyber guardrails removed for authorized defensive work, and Red, which gates access to purpose-trained cyber models. The new model is GPT-5.6-Cyber, trained on top of Sol to handle exploit-chain development, authentication bypass and privilege escalation with far fewer refusals. In OpenAI's internal Advanced Cybersecurity Completion Rate evaluation it answered 95.0% of those requests; GPT-5.6 Sol with safeguards answered 1.5%, and the previous GPT-5.5-Cyber managed 57.3%. On ExploitGym it also beat both predecessors, and OpenAI says it found two previously undisclosed V8 JavaScript engine bugs (fixed as CVE-2026-15903).
The access controls are the story as much as the model. Every Daybreak account now requires a hardware security key, starting September 1, 2026 — today — plus identity verification, monitoring, legal attestations and scope restrictions. OpenAI is also pushing Codex users toward auto-review mode, which gates elevated-permission actions before execution. Under its own Preparedness Framework the model rates High, not Critical. My read: the 95% figure is a refusal-rate number, not an exploit-success number, and the real signal is that OpenAI is comfortable putting frontier cyber tools in approved hands at scale while the industry is still debating what "approved" means after the Hugging Face incident — which, it stresses, GPT-5.6-Cyber was not involved in.
— OpenAI (official) · IT Daily · Help Center
🔗 OpenAI: Expanding Daybreak as the Cyber Defense Window Narrows · IT Daily on GPT-5.6-Cyber and the key requirement · OpenAI Help Center: Daybreak Trusted Access for Cyber
Tesla froze Optimus V3 and is pushing supply to 1,000 units a week in September
Tesla's Optimus moved from demo to production ramp this month. After a JPMorgan analyst visit on August 19, the firm confirmed the third-generation Optimus V3 has passed design freeze with the supply chain largely locked, and that Tesla converted the Fremont Model S/X line — idled since May — into a robot-only line in 46 days, with equipment now being installed. Musk had already laid out the timeline on the Q1 earnings call: scaled production starting late July through August. Supply-chain guidance reportedly pushes capacity to 1,000 units a week in September and 2,000–2,500 a week by year-end, which implies the parts base for roughly 100,000 robots a year, against a Fremont design capacity eventually targeting one million.
The Gen3 hardware is specific: 173 cm tall, 57 kg, roughly ten hours of continuous work, 22 degrees of freedom per hand and 38 across the body. The cost story is the part investors keep circling — about 70% of core components (precision reducers, servo motors, sensors, structural parts) come from Chinese suppliers, and roughly seven in ten of those have passed Tesla audits. Shanghai has already deployed about 50 Optimus units in assembly-line trial runs. My read: the honest question is not whether Optimus works — it does, in narrow tasks — but whether the 2027 external sales target survives the first real production ramp, where yield, not demo quality, decides. Industrial use comes first; Musk says consumer sales are two to three years away.
— Tesla (official, Q1 earnings call) · JPMorgan analyst note · Kalkine Media
🔗 JPMorgan site visit note via Kalkine on the Fremont ramp · 网易 on the August robot monthly report and V3 design freeze · 机构研报 on the 1,000/week supply guidance
IBM open-sourced Granite 4.2, reasoning models trained inside live agent environments
IBM released Granite 4.2 on August 25: three dense models at 3B, 8B and 30B under Apache 2.0, built on roughly 15 trillion pre-training tokens, with a native context of 128K that extends to 512K. The hook is a switchable thinking mode — full chain-of-thought, non-thinking, and a low-effort budget — so one checkpoint can either reason step by step or answer directly, and the 8B and 30B went through an extra "agentic RL" phase where the model actually acts inside sandboxed software-engineering, terminal and web-search environments rather than just predicting tokens. IBM reports SWE-bench Verified of 57.00 for the 30B and 47.67 for the 8B, Terminal-Bench 2.1 of 29.24 and 20.56, and AIME25 of 89.17/86.67/78.33 across the three sizes.
The deployment point matters more than the leaderboard. These are dense, not MoE, so they run on ordinary servers; the reasoning switch is a cost feature, letting companies skip the thinking tokens on the 80% of traffic that never needed them. IBM also trained on 1 trillion tokens of synthetic code from its CodeAlchemy pipeline and shipped two 470M-parameter Granite Speech models. My read: nothing here beats the frontier closed models — IBM doesn't claim it does — but an 8B model scoring in the high 40s on SWE-bench that you can self-host, fine-tune and bill no per-token fee on changes the math for a mid-sized company more than a two-point gain at the top of a leaderboard. The question is whether "good enough and yours" beats "better and rented" for enterprise agent workloads.
— IBM Research (official) · Hugging Face · Unite.AI
🔗 IBM Research: Granite 4.2 brings native reasoning to enterprise agents · Hugging Face: Granite-4.2-8B model card · Unite.AI on the agentic RL pipeline
Google starts removing Google Assistant on September 4 — Gemini takes the phone
Google has told users that it will begin removing Google Assistant from Android phones and tablets on September 4, 2026, rolling out over several weeks. Once the transition reaches a device, there is no way back: saying "Hey Google" or long-pressing the power button will invoke Gemini instead. The shutdown covers Wear OS watches, compatible headphones and earbuds, and Android Auto projected from a phone. Cars with Google built-in (Android Automotive) keep Assistant beyond the deadline, and Google TV and Home speakers follow later on a separate schedule. Assistant launched on May 18, 2016 — a ten-year run that ends with the assistant becoming an agent.
The practical risk is the switchover nobody has to consent to. Google says the process is automatic for eligible devices, and the "switch back to Assistant" toggle disappears once rollout completes. For most users Gemini handles timers, reminders and navigation fine; the gaps show up in the long tail — routines, third-party voice integrations, and the things Assistant could do that Gemini still routes differently. My read: this is the biggest real-world agent migration ever attempted, measured in a billion-plus devices, and the September 4 date is when "assistant" stops being a category and becomes a default. The interesting number to watch is not how many users it reaches, but how many of them notice the difference at all.
— Google (official support notice) · TechRepublic · Deccan Herald
🔗 Google: update on upgrading mobile Assistant devices to Gemini · TechRepublic on who is affected and what to test · Deccan Herald on the September 4 timeline
DeepSeek released V4-Flash-Vision-Exp, its first multimodal V4 model, under MIT
DeepSeek published the weights of DeepSeek-V4-Flash-Vision-Exp on Hugging Face on August 31, ten days after the API opened. It is the first vision model in the V4 family: 305B total parameters with 13B active per token, packed into a 168 GB checkpoint split across 48 safetensors, built by bolting a vision encoder and aligner onto the V4-Flash text architecture (DFlash attention, MoE, Hyper-Connections, DSpark forward path). The license is MIT, and the repo ships a minimal PyTorch inference implementation plus vLLM and SGLang serving instructions.
The benchmark story is mixed, which is refreshing. DeepSeek's own numbers (Harness minimal config, maximum reasoning) show wins over Claude Opus 4.8 on ZeroBench Pass@5 (35.0 vs 34.0) and Agents' Last Exam (27.3 vs 25.7), a win on DeepSWE (59.3 vs 58.0), but losses on ApexBench (36.5 vs 39.4), Chartography (64.3 vs 65.0) and most text-agent tests (NL2Repo 57.7 vs 69.7). Two caveats belong on the record: the Opus comparison uses a superseded 4.8 checkpoint, and all scores are self-reported. My read: the meaningful part is the licensing and the cadence — an MIT 305B multimodal model arriving two days after Tencent's Hy4 preview, and "Exp" tags telling you DeepSeek is testing vision before a wider rollout. The 168 GB checkpoint also quietly raises the bar for who can actually use open weights: the freedom is real, but so is the hardware bill.
— Hugging Face (official model card) · ModelScope · AI Weekly
🔗 Hugging Face: DeepSeek-V4-Flash-Vision-Exp · ModelScope on the weights release · AI Weekly on the benchmark table and the MIT license
OpenAI previewed Private Safety Processing — safety signals without ever reading the data
OpenAI previewed Private Safety Processing on August 19 as the answer to a specific problem: its Zero Data Retention (ZDR) promise says prompts and responses are not retained and never reviewed by personnel, but the most serious risks only show up across multiple interactions. The new system correlates related interactions and returns a narrowly defined signal — the type of risky activity detected — without exposing the underlying content. It works in both modes: customer-controlled infrastructure (ZDR deployments) or OpenAI storage encrypted with keys the customer holds, to which OpenAI personnel have no copy. Early customers shaping the work include Glean, Databricks, Abridge and Microsoft, and OpenAI plans to start rolling it out with a technical white paper in September.
The positioning is pointed. Some providers — Anthropic most visibly — now require retaining customer content for safety monitoring on their most capable models; its risk report says that policy may hurt business if competitors don't follow. OpenAI is betting the opposite: the customer holds the case, the provider holds the alarm. Analysts quoted on the preview (Greyhound, Gartner) make the sharpest point — this is a dispute about where evidence lives, not about privacy versus surveillance, and a system that detects behavior across time has to remember something across time. My read: the September white paper is the actual test, because until someone outside OpenAI can check the claim, "nobody reads your data" is a promise, not a verified property. For regulated industries the practical difference is real either way.
— OpenAI (official) · CSO Online · The Next Web
🔗 OpenAI: Offering Zero Data Retention for frontier models · CSO Online on the signal-based design · The Next Web on the Anthropic contrast
Claude Code weekly limits go up 25% on September 14 — which is 17% less than today
Anthropic's @ClaudeDevs account announced on August 30 that standard weekly limits in Claude Code will permanently rise 25% above baseline for Pro, Max, Team and seat-based Enterprise plans starting September 14. The catch is in the reference points. The current temporary +50% boost stays in place until September 13, so normalizing the baseline to 100: today's limit is 150, and from September 14 it becomes 125. Anthropic itself acknowledged the optics in a follow-up — a 25% increase over the old baseline is a 17% reduction compared with what users have today — and said it is preparing usage-visibility improvements.
Two caveats belong in the summary. First, the support documentation still lists the +50% promotion as ending August 31, so the September 14 schedule currently rests on the official X announcement alone. Second, the change covers weekly limits only — nothing about the five-hour session cap, pricing or plan structure. My read: this is the first concrete sign of the "usage-limit era" for coding agents — Anthropic raised limits twice in a summer as a growth lever, and now it is normalizing them upward permanently while the temporary boost quietly expires. The 17% framing is the part worth remembering next time a lab announces a "raise": always ask, compared to when?
— Anthropic @ClaudeDevs (official X) · UsingClaude · AI Catchup
🔗 @ClaudeDevs announcement on X · UsingClaude on the 25% vs 17% math · AI Catchup on the September 14 schedule
Top comments (0)