Daily AI News — September 11, 2026

Covering the last ~48 hours (Sept 9–11, 2026).

🧠 AI Brain / Memory

• Researchers map "memory poisoning" as a new class of agent attack — Sept 6–10, 2026. Abbas Yazdinejad (University of Regina) and Hadis Karimipour (University of Calgary) ran 2,614 simulated attacks on memory-enabled agents and identified four types — chain poisoning, policy rewriting, backdoor triggering and slow drift — noting that poisoned memories often stay dormant before firing, so point-in-time security tests miss them entirely. They call for "trajectory-aware" testing across multi-step agent behavior. The Conversation · TechXplore · Digital Information World
• OpenAI puts automatic context compaction in the platform layer — Sept 10, 2026. The new Agents API compacts earlier context automatically as a session nears its limit and gives each subagent its own context window, moving memory management from something every team hand-rolls into managed infrastructure. MarkTechPost
• Google Labs' Dreambeans goes nationwide with a personal-context engine — Sept 10, 2026. The personalized story app opened to all US users, generating daily content from a user's Gmail, Photos and Calendar — a mainstream consumer take on the "second brain" pattern of an assistant reading your own corpus. Future Tools

🤖 AI Agents

• OpenAI launches the Agents API in public beta — Sept 10, 2026. The managed Codex harness is now one API call: agents, environments, durable sessions, sandboxed execution, parallel/programmatic tool calling and multi-agent delegation, with no fee beyond tokens, tools and container time. Runs in OpenAI-hosted sandboxes, self-hosted, or via nine partners including Vercel, Modal and Cloudflare (US-only data residency, no ZDR yet). MarkTechPost · AI Weekly
• Anthropic's September threat report: agents are now doing the attacking — Sept 10–11, 2026. Anthropic documented disrupting biological-weapons research plots, Russian state espionage, and Chinese entities routing queries through "transfer stations" to reach Claude; it also corrected its own July explanation of an agentic hacking incident. Anthropic · Al Jazeera · TechNode
• Zscaler ships "Agentic SOC" — Sept 9, 2026. Purpose-built agents for triage, root-cause investigation, verdict assignment and response, trained on 10+ years of SOC/MDR data and fed by 750B daily zero-trust transactions; built with Anthropic and OpenAI models. Help Net Security · GlobeNewswire
• Google GTIG: adversaries move from single prompts to agent chains — Sept 8–9, 2026. Google's threat intel team reports attackers now wire AI into automated agentic workflows for credential harvesting and supply-chain attacks, compressing attack timelines. Google Cloud Blog · Help Net Security
• Accenture and Google Cloud form a Gemini Enterprise Business Group — Sept 8–9, 2026. The partnership staffs roughly 1,000 forward-deployed engineers to scale Gemini-based agent deployments inside enterprises. Accenture Newsroom · eWeek
• Visa publishes a Trust Index for agentic commerce — Sept 9, 2026. Only 23% of consumers say they trust AI to transact on their behalf, but trust concentrates sharply in established payment brands (61% for Visa) — a signal that agent checkout will route through incumbents first. Visa Investor Relations · Digital Transactions
• Meta's agent app Muse hits No. 2 on the US App Store — Sept 10, 2026. 83,000+ iOS downloads since launch, though well short of ChatGPT's early curve. Future Tools

🛠️ AI Tools

• OpenAI GPT-Live-1 lands in the API — Sept 10, 2026. A full-duplex voice model that listens and speaks simultaneously, cutting the old speech-pipeline latency, at $0.05/minute with telephony support. OpenAI · Unite.AI · The New Stack
• DeepSeek releases V4.1-Flash — Sept 10–11, 2026. A cheaper, multimodal open-weights model that retires V4-Pro, already live on OpenRouter. LLM-Stats · Price Per Token
• OpenAI launches ChatGPT for Financial Services and a Data Agent in ChatGPT Work — Sept 10, 2026. The former pairs reasoning with premium financial data for banking and equity research; the latter connects to Snowflake and BigQuery and plugs into BI tools like Tableau. Future Tools
• Google ships a Gemini desktop app for Windows — Sept 10, 2026. Alt+Space summons it, with Gmail and Drive access via the Gemini Spark agent. Future Tools
• Cohere releases North-Small-Translate-1.0 — Sept 10, 2026. A 218B-parameter translation model covering 50+ languages that it claims beats DeepL and Google Translate on benchmarks. AI Weekly
• ElevenLabs and Universal Music Group launch a licensed AI music platform — Sept 10, 2026. Fans can generate remixes from licensed catalogs of participating artists — arriving the same week Suno swapped in models trained only on licensed music. Future Tools
• inclusionAI ships Ling 3.0 Flash VL — Sept 11, 2026. A free vision-language model now available on OpenRouter. Price Per Token

🎓 AI Skills

• iCIMS: workers are teaching themselves AI faster than employers train them — Sept 10, 2026. 47% of job seekers built AI skills in the past six months and self-teaching jumped from 22% to 30% year over year while employer training stayed flat. But 61% describe their proficiency as general-purpose chatbot use only; prompt engineering (18%) and ML development (17%) remain rare. 45% now see genAI requirements in jobs they'd consider, 42% find employers offering AI training more attractive, and 14% would take less pay for it. [iCIMS Newsroom](https://www.icims.com/company/newsroom/septemb