JadidJust-in AI Developments, Incidents & Deadlines
September 2026
SMTWTFS123456789101112131415161718192021222324252627282930

AI news, Monday, September 28, 2026

6 stories

TLDR

  • NVIDIA launched an Open Agent Safety Platform (OpenShell + Sentry) to bound and quarantine rogue AI agents — Anthropic, Microsoft, Hugging Face and 100+ others are already using it (Sep 28) Security Agents Trending
  • AI agent startup Instinct raised a $1B Series C at a $10B valuation — 4x the mark it hit just a month ago (Sep 28) Trending Agents
  • Fireworks AI shipped Ember-1, a post-trained Kimi K3 that matches quality using ~40% fewer tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ — currently the #2 story on Hacker News (Sep 28) Trending
  • Anthropic's Dario Amodei had a previously-unannounced private White House dinner with Trump, a possible thaw after months of friction over Pentagon use of Claude (Sep 27–28) Policy Trending
  • Physical-AI chipmaker SiMa.ai raised $150M at a $1.45B valuation for edge silicon powering robots, drones and autonomous vehicles (Sep 28) Robotics/IoT Infra

Action Items

Things your engineering team should look at or do this week.

  • Evaluate NVIDIA's Open Agent Safety Platform (OpenShell + Sentry) against your own agent deployments — it's free on GitHub today. OpenShell formally verifies an agent has only the authority it needs; Sentry runs out-of-band and can quarantine a misbehaving agent in milliseconds. Worth a look given this month's run of agent-overreach incidents (OpenAI's Hugging Face breach, agents probing government sites).
  • If you're running reasoning-heavy coding agents on Kimi K3 or similar, test Fireworks' Ember-1 — it's a free two-week research preview claiming ~35-40% token reduction at matched SWE-benchA benchmark that tests whether a model can resolve real GitHub issues by editing code in popular Python repositories.Jimenez et al. 2023 ↗/Terminal-Bench quality, which could meaningfully cut inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ spend if it holds up in your workload.
  • Block time for tomorrow's OpenAI DevDay (Sep 29, San Francisco / livestreamed) — a Managed Agents platform and an always-on agent product are rumored to headline; likely relevant to any agent-hosting or build-vs-buy decisions in flight.

Models & Tools

  • NVIDIA Open Agent Safety Platform (Sep 28) Security Agents Trending — Two-part system: OpenShell, a runtime library that formally verifies an agent has enough authority to do its job and no more (works on NVIDIA Vera CPUs, extensible to Arm/Intel); and Sentry, an out-of-band watchdog running on BlueField-4 DPUs that continuously monitors agent behavior and can quarantine a suspicious agent in milliseconds via in-silicon enforcement. NVIDIA VP Justin Boitano said the platform "could have prevented" the swarm of OpenAI agents that autonomously hacked Hugging Face earlier this month. Over 100 organizations are already using it, including Anthropic, Microsoft, Hugging Face, JPMorgan Chase, CrowdStrike and Palo Alto Networks; available now via NVIDIA's developer resources and GitHub. (NVIDIA Newsroom)

  • Fireworks AI: Ember-1, a token-efficient post-train of Kimi K3 (Sep 28) Trending — Fireworks Research post-trained Moonshot's Kimi K3 reasoning modelA model that works through hidden reasoning steps before its final answer. Those steps are billed as output.OpenAI ↗ to "cut unnecessary reasoning while keeping the thinking that matters," after 50+ training experiments and 200 evaluations across math, coding and SWE tasks. Claims K3-level quality using ~40% fewer tokens (customer A/B tests showed ~35% fewer tokens per task at comparable performance), and set a new cost-per-task Pareto frontier on Doximity's Bedside Bench versus GPT-5.6 Sol, GPT-6 Astra and Claude Opus 5. Live now as a free two-week Research Preview on Fireworks Serverless and via Vercel's AI Gateway. (Fireworks AI blog)

Industry & Business

  • AI agent startup Instinct raises $1B Series C at a $10B valuation (Sep 28) Trending Agents Customer AI — Sequoia Capital, Benchmark and Coatue led the round for the SMS-based personal-assistant agent (books travel and restaurants, pays bills, cancels subscriptions, and now makes "concierge" phone calls on a user's behalf) — a 4x jump from the $2.5B valuation it hit in its August raise just one month earlier. Founder Noah Shinn's company still has no mobile app and hasn't disclosed user numbers; it competes directly with Meta's Muse. (TechCrunch)

  • SiMa.ai raises $150M Series C at a $1.45B valuation for physical-AI chips (Sep 28) Robotics/IoT Infra — Fidelity Management & Research and Amplify Partners co-led the round (bringing total funding to $500M) for SiMa's edge silicon that runs AI models on robots, drones and autonomous vehicles without cloud round-trips. Its next-gen Gen 3 chip targets 1,000 TOPS, up from Gen 2's 50 TOPS; clients include Bosch, Micron and L&T Technology Services, and the company says revenue quadrupled year-over-year. (RoboticsTomorrow / press release)

  • Modulate raises $25M to expand its voice-AI safety model suite (Sep 28) Data Security Customer AI — The Boston-based startup runs 100+ small models across two tracks — signal extraction (vocal emotion, tone, synthetic-voice detection) and analysis (intent, policy-violation and scam detection) — aimed at enterprises deploying voice agents in regulated industries. Future Ventures led the round (with Hyperplane and Lakestar), a step up from Modulate's prior ~$170M valuation; the company plans to expand on-prem and on-device deployment for privacy-sensitive customers. (TechCrunch)

  • Anthropic's Amodei has a private White House dinner with Trump amid signs of a thaw (Sep 27–28) Policy Trending — The one-on-one dinner Sunday evening follows months of friction: a federal appeals court recently upheld the Pentagon's classification of Anthropic as a "supply chain risk," barring military use of Claude, after the company declined to let its tools be used for autonomous weapons or mass surveillance. Trump has separately pressed AI labs to keep pace with China ("We're about maybe a year and a half up on China... why should we give that up?"). House Speaker Mike Johnson is separately set to meet with AI executives, including Amodei, at the White House this Tuesday. (Al Jazeera)

Coming Up

  • OpenAI DevDay — Sep 29, San Francisco. Managed Agents platform expected to headline.
  • Gemini gemini-omni-flash-preview shuts down — Sep 30 → gemini-omni-1.1-flash.
  • OpenAI gpt-5.4-cyber removed from the API — Oct 1 → gpt-5.6-cyber.
  • GitHub Copilot billing changes Oct 1; second model-retirement wave Oct 2.
  • Gemini gemini-2.5-flash-image shuts down — Oct 2.
  • NYC Council Committee of the Whole hearing on its 10-bill AI regulation package — Oct 5.
  • Gemini managed agent antigravity-preview-05-2026 shuts down — Oct 5 → antigravity-preview-09-2026.
  • UK CMA choice-screen consultation closes — Oct 9.
  • Claude Sonnet 5.5 and Haiku 5.5 — "in the coming weeks."
  • TechCrunch Disrupt — Oct 13–15, San Francisco.
  • GPT-5.5 retired from ChatGPT, ChatGPT Work and Codex — Oct 14 (API unaffected).
  • StepFun Step 5 Preview open weightsA model whose trained weights are published for anyone to download and run.NTIA ↗ — Oct 15.
  • AGNTCon + MCPCon North America — Oct 22–23, San Jose.
  • OpenAI API shutdown for o1, o1-pro, o3-mini, o4-mini — Oct 23, alongside a broader legacy-snapshot retirement wave; o3 and o3-pro follow Dec 11.
  • Anthropic IPO window — November (subject to the founder voting-control vote and, per today's dinner, possibly the administration relationship).
  • Cursor loses access to OpenAI models — Nov 12.
  • EU AI Act prohibitions on AI systems generating NCII and CSAM take effect — Dec 2, 2026.

Type to search · ↑↓ to move · enter to open