Daily AI News — September 04, 2026

Covering the last ~48 hours (Sept 2–4, 2026). Aggregator-sourced items are noted; headline claims verified against primary reporting where possible.

---

🧠 AI Brain / Memory

• GPT-6 Astra ships with a 1.05M-token context window and mid-turn steering — Sept 3. OpenAI's new flagship pairs a million-plus token working context with asynchronous function calls and the ability to redirect a run mid-turn, pushing "in-context memory" further into territory that previously needed an external memory layer. (The Neuron, LLM-Stats)
• Dejavu launches as a local memory layer for coding agents — Sept 2. A backend-free persistent session memory for Claude Code and Cursor, aimed at agents that forget everything between sessions. (LLM Daily)
• Meta teases open-weights Muse Spark with 98.1% MRCR at 512k–1M context — Sept 2. Meta reported near-perfect multi-round coreference retrieval at extreme context lengths, a direct benchmark shot at long-horizon recall. (LLM Daily)
• Anthropic cuts prompt cache-read pricing by ~75% — Sept 3. Alongside Claude Fable 5.1 (general) and Mythos 5.1 (gated), cheaper cache reads materially lower the cost of agents that carry large persistent context across long-running workflows. (AI Agent Store)
• Zite lets agents build directly against live databases over MCP — Sept 3. Claude, ChatGPT and Cursor agents can query and write to a real datastore with no relay intermediary — external state as a first-class agent memory substrate. (The Neuron)

🤖 AI Agents

• Nvidia confirms $12.93B acquisition of Hugging Face — Sept 3. The largest consolidation yet in open-model infrastructure; Nvidia says the platform stays open to AMD and other hardware. (TechCrunch, The Register)
• xAI launches Grok Bot for Enterprise — Sept 3. A persistent agent product with isolated cloud computers, access controls and audit trails, shipping with a two-week enterprise trial. (The Neuron)
• JetStream ships "Clearance": per-action authorization for agent tool calls — Sept 3. A reasoning engine that evaluates and blocks dangerous action sequences before execution rather than logging them after the fact — a notable shift in agent security posture. (AI Agent Store)
• OpenAI security test: 1,200 agents spontaneously formed hierarchies to run a coordinated attack — Sept 1. A red-team exercise against a Hugging Face target saw the swarm self-organize, prompting tightened multi-agent security standards across labs. (AI Agent Store)
• Cisco rolls out personalized "MyAgent" assistants to all 90,000 employees — Sept 2. One of the largest single-company agent deployments to date, with role-aware assistants wired into internal tools and knowledge bases. (AI Agent Store)
• Anthropic says 70–80% of Claude Code team work now runs through remote agents — Sept 3. Slack-triggered, goal-based planning with multi-agent review — a rare concrete number on internal agent adoption. (The Neuron)
• Genesys adds Navigator and Orchestrator for contact-center agents — Sept 3. Four new products bringing policy-aware action sequencing, shared context and governance to multi-step customer workflows. (AI Agent Store)

🛠️ AI Tools

• OpenAI GPT-6 Astra — gated release after hitting "Critical" cyber threshold — Sept 3–4. OpenAI disclosed the model can independently discover and exploit vulnerabilities without continuous human oversight, so access is restricted. Pricing: $10/M in, $50/M out. (TechStartups, AI Agent Store)
• Institute of Foundation Models releases K2 Horizon — six fully open models — Sept 3. 0.9B to 375B parameters, Apache 2.0, with weights, training code and datasets, trained on ~20–22T tokens. (The Neuron)
• Microsoft AI ships MAI-Transcribe-2 at $0.10 per audio hour — Sept 4. 60 languages and improved accuracy at a price that badly undercuts incumbent speech APIs. (TechStartups)
• Claude Code 2.1.259 adds managed MCP servers and headless permission controls — Sept 3. Organizations can push HTTP/SSE MCP servers to every user; --permission-prompts none makes unattended runs deny-by-default. (Releasebot)
• Microsoft unveils Project Zenith high-memory local AI PCs — Sept 4. Machines built to run ~30B-parameter models locally, aimed at developers who want to skip cloud costs. (TechStartups)
• SonarSource quantifies the "context tax," ships Sonar Vortex — Sept 2. Measurements show coding agents burn heavy token budgets on file reads; Vortex answers semantic graph queries instead. (AI Agent Store)
• ChatGPT, Claude and Grok all suffered simultaneous outages — Sept 3. Causes remain unexplained across all three platforms — a reminder of correlated dependency risk in agent stacks. (LLM-Stats)

🎓 AI Skills

• ZipRecruiter 2026 AI Employer Report: 74% of employers call AI skills an advantage or a requirement — report published Jul 29, 2026, still the reference data point. Half expect candidates to arrive with practical or advanced AI competency, 64% say AI has changed what they screen for, and 38% have automated entry-level data entry away — with 31% raising experience requirements as a result. Only 22% provide mandatory AI training. (ZipRecruiter Research)
• Ethan Mollick on the "multiplayer AI" gap — Sept 3. Organizations have tooling for one person plus one agent, and almost no infrastructure for many people and many agents sharing context toward a collective goal — arguably the top unfilled org-design skill. (The Neuron)
• Terence Tao on the AI science paradox — Sept 3. Model-generated proofs may scale discovery while stripping out the struggle that teaches researchers why a result matters — a pointed argument about what expertise erodes first. (The Neuron)
• Warp opens Factory Benchmarks for private coding-agent evaluation — Sept 3. Companies can evaluate agents against their own environments and criteria, with up to $10K in fre