Daily AI News — September 12, 2026

Covering the last ~48 hours (Sep 10–12, 2026).

🧠 AI Brain / Memory

• Meta details its "Organizational Second Brain" agent architecture — Sep 9 (Meta post Sep 2). Meta separates what an agent knows (version-controlled knowledge files) from how it reasons (explicit recipes), so expert corrections compile into verified, regression-tested edits with no model retraining — cutting compliance assessments from days to minutes. (InfoQ, Engineering at Meta)
• Cursor's Projects beta puts memory in a coordinator agent — Sep 10. A delegating coordinator "directs other agents and keeps the context," making persistent shared context — not raw model capability — the unit of multi-agent dev work. (AI Weekly)
• OpenAI's Agents API ships managed sessions as a first-class primitive — Sep 10. The beta "manages sessions, tool coordination and recovery," meaning session/memory state is now vendor-managed infrastructure rather than something every team rebuilds. (OpenAI Developers, MarkTechPost)
• Context: the memory-layer market keeps consolidating — Mem0's State of AI Agent Memory 2026 benchmarks remains the reference point for where persistent-context tooling stands this year. (Mem0)

🤖 AI Agents

• Salesforce launches seven named, "job-ready" Agentforce agents — Sep 12. Casey (service), Paige (HR/IT), Carter (commerce) and others ship as pre-scoped roles rather than toolkits; Salesforce claims 7 billion Agentic Work Units delivered to date. (Salesforce)
• ⚠️ Attackers ran hundreds of AI agents against PaperCut — Sep 11. A Russian-speaking threat actor used agents built on OpenAI Codex and DeepSeek to exploit PaperCut NG/MF, compromising at least 440 instances across 48 countries — agentic attack tooling at real scale. (Help Net Security)
• ⚠️ Critical RCE patched in Google's Agent Development Kit — Sep 10. CVE-2026-79696 (CVSS 10.0) allowed unauthenticated remote code execution via crafted test replays; anyone on ADK should verify their version. (OffSeq Radar)
• Enterprises are overconfident on agent safety — Sep 11. A Harness survey found 77% claim an agent inventory but only 44% actually run discovery — trust is outpacing controls. (AI Agent Store)
• Zscaler debuts "Agentic SOC"; Visa publishes an agentic-commerce Trust Index — Sep 10. Zscaler embeds triage/containment agents in security ops, while Visa's index shows consumers trust established payment brands over generic AI tools for agent-initiated purchases — payments as a control point. (AI Agent Store)
• Verticals keep filling in — Sep 10–11. Google Cloud launched banking agents with Deutsche Bank as proving ground; FINTRX shipped "Fin" for private-wealth monitoring; Sixfold launched an AI underwriter; Sidetrade released a sovereign CFO agent framework. (Yahoo Finance, Agentic.ai)

🛠️ AI Tools

• Sakana AI releases Fugu Max, a learned orchestrator — Sep 11. Routes queries across open-weight models at a claimed 40–60% below Sonnet 5 and GPT-5.6 Terra pricing — routing as the cost lever. (Sakana AI)
• Cognition ships SWE-2 coding agent — Sep 10. Scores 50.0% on FrontierCode 1.1 Main while running ~64% cheaper than comparable models. (Cognition)
• OpenAI's GPT-Live-1 brings full-duplex voice to the API — Sep 10. Interruptible, natural-turn conversation at $0.05/minute, materially changing what voice agents can feel like. (AI Weekly)
• OpenAI Data Agent lands in ChatGPT Work — Sep 10. Connects business data sources and builds interactive dashboards from analytics questions. (Blockchain News)
• Cohere's North Small Translate + DeepSeek-V4.1-Flash — Sep 10. A 218B-parameter, 50-language MoE translation model claimed to beat DeepL and Google; DeepSeek shipped a new Flash tier the same day. (AI Weekly, AI Release Tracker)
• ⚠️ Coding-sandbox escapes disclosed across Claude Code, Codex and Cursor — Sep 11. Stealth startup Accomplish claims leaky sandboxes in all three; Anthropic reportedly took ~50 days to patch vs ~a week for competitors. (Upstarts Media)

🎓 AI Skills

• iCIMS: workers are teaching themselves AI faster than employers train them — Sep 10. Self-teaching jumped 22% → 30% in a year and 47% of job seekers built AI skills in the last six months, while AI postings are still just 4% of U.S. hiring demand. (iCIMS, PR Newswire)
• The confidence gap is the real story — Sep 10. 60% of job seekers feel prepared for workplace AI, but 61% stop at general tools like ChatGPT and only 18% report prompt-engineering skill — surface familiarity, not agent-building capability. (iCIMS)
• AI training is now a recruiting lever — Sep 10. 42% of candidates say an employer offering AI training is more attractive than competitors; finance leads AI skill saturation, healthcare is the fastest-emerging. (iCIMS)
• CyBe: a 14.5-point AI readiness gap inside the same companies — Sep 10. Senior leaders at one organization disagree sharply about readiness for autonomous agents — an internal alignment problem as much as a skills one. (ANI)
• Skills-trend baseline — Sep 2026. The Bipartisan Policy Center's monthly skills dashboard tracks how AI requirements are moving across occupations. (BPC)

---

Signal of the day

Agent infrastructure went managed (OpenAI Agents API, Cursor coordinator, Agentforce roles) in the same 48 hours that agent security broke badly — a CVSS 10.0 RCE in Google's ADK, 440 PaperCut instances hit by attacker-run agent fleets, and sandbox escapes across all three major coding agents. Capability is shipping faster than the controls around it, and the enterprise survey data says teams don't know it yet.