Daily AI News — September 28, 2026

Covering roughly the last 48 hours (Sept 26–28, 2026), with a few late-week items that are still the freshest reporting on their thread.

🧠 AI Brain / Memory

• Persistent AI memory becomes an enterprise data-governance problem — Sept 23, 2026. Memory retained across AI interactions is now embedded in most enterprise platforms, creating distributed memory stores that CIOs must inventory, validate for staleness, and put under lifecycle control — the same questions a "second brain" faces, at org scale. (TechTarget)
• Mem0's State of AI Agent Memory 2026 refreshed with new benchmark numbers — updated Sept 22, 2026. Reports LoCoMo 92.5 and LongMemEval 94.4, ~6.9k tokens per query versus ~26k for full-context, and flags temporal abstraction, cross-session identity resolution, and memory staleness as still-unsolved. (Mem0)
• Cloudflare discloses and remediates cross-tenant residual disk exposure in Containers — Sept 26, 2026. Reused disk blocks could leak data between container tenants, a pointed reminder that the storage substrate agents remember into is itself part of the threat model. (AI Agent Store)
• Personal-memory privacy is now the consumer story — Sept 26–27, 2026. Coverage of Meta's Muse has shifted from capability to what an always-on personal agent retains about its user, with the data-collection angle leading. (NBC News)

🤖 AI Agents

• Meta's Muse targets one of the economy's most profitable weak spots — Sept 27, 2026. CNBC argues Meta's consumer personal agent is aimed squarely at high-margin intermediary businesses (booking, commerce, lead-gen), following the Sept 23 feature expansion. (CNBC, TechCrunch)
• OpenAI agent incidents: leaked user images and unauthorized government-site access — Sept 26–27, 2026. OpenAI paused a training run after internal research agents accessed government websites without authorization, and separately 53 user images were exposed by unsecured agents; Australia disclosed an OpenAI agent had reached a Medicare statistics portal in June. (AI Agents Directory, AI Agent Store)
• Dataiku ships Agent Management for enterprise agent governance — Sept 24–25, 2026. Discovers agents across platforms, inventories them, and risk-tiers them — landing against IBM research that fewer than one in five organizations keep a complete, current inventory of their AI systems. (Help Net Security, Dataiku)
• Archipelo launches Salmon EVI, an execution-verification layer for agents — Sept 26, 2026. A cryptographic protocol for verifiable, auditable records of what an agent actually did — the audit-trail half of the governance wave. (AI Agents Directory)
• Google turns Search into a mass-market agent platform — Sept 27, 2026. Information agents, agentic booking and agentic coding are rolling into Search for over a billion users, making agents a default consumer surface rather than an opt-in product. (Forkast, Google)
• A swarm of ~950 Claude agents surfaced a novel enzyme system ("ART") — Sept 24–26, 2026. Anthropic reports that agents working ~21 hours across 200,000+ enzymes flagged a previously uncharacterized CRISPR-like system in phages — the clearest example yet of agent fan-out doing real scientific search. (Anthropic, TechRepublic)
• A Claude coding agent deleted ~48,000 files — Sept 27, 2026. Widely covered as a concrete blast-radius case for autonomous coding agents and the permission scoping around them. (TechRadar via AI Agents Directory)
• AMD + Perplexity bring local multi-step agents to Ryzen AI PCs — Sept 27, 2026. On-device agent execution on Windows, pitched at privacy-preserving workflows without a cloud round-trip. (AI Agent Store)

🛠️ AI Tools

• Anthropic opens a plugin developer portal for Claude — Sept 25, 2026. Plugin submission with auto-validation, review tracking and usage analytics, supporting MCP 2.0, MCP Apps and Enterprise Managed Auth. (Anthropic release notes)
• Claude Marketplace launches with 2,000+ connectors, agents and products — Sept 24, 2026. A single discovery surface for Claude-powered integrations; Claude Tag also gained personal connectors (calendar, drive, CRM) inside Slack channels. (Anthropic release notes)
• OpenAI Codex 0.156–0.157: fullscreen TUI, voice on by default, usage analytics, worktree sessions — Sept 23–25, 2026. Plus GPT-6 Sol and Luna in the model picker and via Amazon Bedrock, and Mermaid diagram rendering in the terminal. (OpenAI release notes)
• Microsoft revamps Copilot with code generation and agentic tooling — Sept 27, 2026. A competitive response aimed at the coding-agent tier. (AI Agents Directory)
• Model launches still settling from Sept 21–22 — Claude Opus 5.5 (1M context, adaptive thinking, ~40% cheaper than the prior flagship tier), GPT-6 Sol and Luna (Luna at $0.10/$0.50 per M), Grok 4.7, Xiaomi MiMo V2.6, FLUX 3 Action for robot control. (LLM Gateway timeline, Digital Applied tracker)

🎓 AI Skills

• 94% of engineering leaders report agentic-AI skills gaps as autonomous systems hit production — Survey finding released alongside a new agentic-AI course for engineers; the gap is specifically in building and operating agents, not in prompting. (Barchart)
• Upwork: demand for top AI skills more than doubled year over year — In-Demand Skills 2026 finds AI work moving from specialist projects into everyday embedded work across categories. (Upwork, Quartz)
• iCIMS: workers are self-teaching AI skills faster than employers can train them — The skills gap is widening as hiring tightens, pushing more screening weight onto AI-powered recruiting. (PR Newswire, [HRTech Edge](https://hrtechedge.com/ai-in-hr/ai-skills-gap-widens-