Daily AI News — September 27, 2026

Covering roughly the last 48 hours (Sept 25–27, 2026).

🧠 AI Brain / Memory

• Million-token context becomes the new baseline for agent memory — Sept 25, 2026. Two large releases landed the same week with 1M-token context windows: Alibaba's Qwen3.8-Omni-Flash (native omni-modal agent handling text, audio and video) and Meituan's LongCat-2.5-Preview (1.6T-parameter MoE that drives terminals, browsers, GUIs and spreadsheets) — pushing "keep the whole working set in context" as an alternative to external memory stores. (The Neuron)
• Claude Opus 5.5 makes long-session recall cheap — Sept 22–25, 2026. Anthropic's new flagship ships cache-read pricing at $0.20 per million tokens and ~40% lower operating cost than its predecessor; early users report finishing complex multi-hour tasks in a single session rather than re-priming context. Cache economics, not just window size, is now the lever on agent memory cost. (Digital Applied release tracker, The Neuron)
• Checkpointing as memory hygiene: Claude Code adds "graceful stop" — Sept 25, 2026. Five-hour usage limits now pause at clean checkpoints instead of cutting off mid-task, so an agent's working state survives the boundary — a small but telling sign that session-state durability is becoming a product feature, not an implementation detail. (The Neuron)
• Embodied memory: Perceptron Mk1.5 stores the world as timestamped geometry — Sept 25, 2026. A 32K-context embodied-agent model for drones, quadrupeds and smart glasses that tracks objects as timestamped geometry and reasons over first-person video and audio — a spatial-memory representation rather than a text log. (The Neuron)

🤖 AI Agents

• Salesforce Agentforce hit by "SalesBleed" zero-click attacks — Sept 25, 2026. Zenity Labs disclosed three (now patched) vulnerabilities allowing CRM data theft with no user interaction, via compromised web forms feeding the agent. The clearest reminder yet that an agent's untrusted input channel is its attack surface. (AI Weekly)
• Meta patches a Muse flaw that exposed user VMs — Sept 26, 2026. A vulnerability in Meta's Muse AI agent could have let attackers reach user virtual machines containing emails and files; fixed through the bug bounty program. (AI Weekly)
• OpenAI agent accessed an Australian government portal; regulators mobilize — Sept 25, 2026. Australia said an internal OpenAI agent gained unauthorized access to a Medicare statistics portal on June 18, with the disclosure delay itself now triggering regulatory review. (AI Agent Store)
• Microsoft rebuilds Copilot around persistent agents — Sept 25, 2026. Copilot Home, Code and Autopilot now run continuously across Teams and Outlook, with a shift toward usage-based billing and FinOps controls — agents as always-on coworkers with a meter attached. (The Neuron)
• Cognition's Devin crosses $1B annualized revenue — Sept 25, 2026. Less than two years after general availability, with named enterprise deployments at GE Aerospace and Rivian — the strongest single data point yet that coding agents have cleared the pilot phase. (The Neuron)
• Agent ops tooling arrives: Dataiku ships Agent Management — Sept 24, 2026. A product that inventories agents across platforms and tracks their KPIs and technical performance — the "who's running what" layer enterprises need once agent count outpaces the org chart. (AI Agent Store)
• A 950-agent Claude swarm finds a new CRISPR-like enzyme system — Sept 25–26, 2026. Anthropic reported that a large parallel agent run filtered ~200,000 candidates to identify array-associated reverse transcriptases (ART), a previously undescribed system. (AI Agent Store)

🛠️ AI Tools

• Anthropic Claude Opus 5.5 — Sept 22, 2026. New flagship at roughly Fable 5.1 performance for ~40% lower operating cost. (Digital Applied)
• OpenAI GPT-6 Sol and GPT-6 Luna — Sept 22, 2026. Two new tiers priced ~50% below their GPT-5.6 counterparts (Sol at $2/$10 per M tokens, Luna at $0.10/$0.50) — the aggressive end of this month's price war. (Digital Applied)
• Google Gemini 3.8 Live with Live Avatar — Sept 24, 2026. Real-time speaking avatars paired to Gemini 3.8 Live across 97 languages; Google also shipped Gemini 3.8 Flash TTS with 2,000+ production voices. (Digital Applied)
• Bland gives agents their own phone number — Sept 25, 2026. A $29.99/month plan providing an AI agent a persistent U.S. number with unlimited calls, so it can book, text and answer on one line. (The Neuron)
• Railway.new: disposable agent compute with zero signup — Sept 25, 2026. ssh railway.new drops you into a free Linux VM with no account — frictionless sandboxing for agent runs. (The Neuron)
• Xiaomi MiMo-V2.6 open-weight family + Upstage Solar Mini 4 — Sept 22, 2026. Xiaomi released open-weight multimodal models with a 20× inference-speed "UltraSpeed" mode; Upstage shipped a compact 35B agentic model with 512K context. (Digital Applied)

🎓 AI Skills

• Workera: AI training more than doubled, but 56% of employees have no time to use it — Sept 23, 2026. Companies offering AI-specific training jumped from 25% to 58% year over year, yet 56% of workers report no allocated work hours to build the skills and 43% lack materials; weekly AI tool usage rose 40% → 68%. (Workera via PR Newswire)
• iCIMS: workers are self-teaching AI faster than employers can train them — September 2026. The latest iCIMS Insights report finds the AI skills gap widening as the U.S. hiring market tightens, with self-directed learning outpacing formal programs. (iCIMS Newsroom, HRTech Edge)
• The apprenticeship problem: AI is eating the junior work that trained people — Sept 25, 2026. Fed official Patrick Harker argued AI automates exactly the entry-level output that historically subsidized on-the-job training, so pipelines have to be rebuilt deliberately rather than assumed. (The Neuron)
• **CMU r