|
$ ai-news --date 2026-09-11 | mail
AI News — September 11, 2026
// the most viral AI topics of the last 24 hours, curated for your stack
$ cat top-story.md
Jacob Coxon, who spent three years training frontier models first at OpenAI and then Anthropic, resigned this week and published a seven-part X thread — nearly 76 million views overnight — warning that both companies are "racing straight to self-improving superintelligence and gambling with our lives." The notable twist: Anthropic's own alignment science lead, Evan Hubinger, publicly backed him instead of distancing the company, saying "I personally think it is >10% [chance of AI killing everyone] within the next decade... we do not yet have a plan to solve alignment for superintelligence." The story dominated Hacker News and mainstream coverage alike, alongside skepticism from commenters questioning how a previously unknown account got such instant reach.
Why it matters: This isn't outside criticism of the lab you build on daily — it's Anthropic's own alignment lead agreeing there's no plan yet to solve alignment for the systems it's racing to ship.
// source: NBC News · additional sources: Washington Post · Hacker News
## 📦 Models & Releases
SWE-2, Cognition's new coding-agent model for Devin, is post-trained from Moonshot's open-weight 2.8T-parameter Kimi K3 and scores 50.0% on FrontierCode 1.1 — within a point of Claude Fable 5.1 at 64% lower cost — while needing a median of just 18 steps to make its first real edit versus 48 for SWE-1.7. Why it matters: A concrete data point on how far RL post-training can push an open-weight MoE base, useful if you're weighing self-hostable coding backbones against frontier-API pricing. // Cognition · Hacker News (373 pts)
One week after launch, demand for GPT-6 Astra pushed OpenAI to pause new sign-ups and upgrades for its $200/month Pro tier, calling the demand "unprecedented"; existing subscribers are unaffected, and Plus, Go and API access remain open. Why it matters: If you wanted to trial Astra via Pro to benchmark against your own stack, the API remains the more reliable path in for now. // TechCrunch
## 🛠️ Tools & Repos
This Show-HN CLI needs no install, account or telemetry: it reads configs and directories to list every agent harness, MCP server, plugin and AI extension present on a box, flags what each can touch, and prints plain-language remediation steps, with a --json mode for fleet-wide drift scanning. Why it matters: With agent harnesses and MCP servers multiplying fast, a read-only audit is a sane first step before trusting a new one with filesystem or network access on your own machines. // Hacker News (Show HN)
CVE-2026-82533 (9.4/10) lets a sandboxed DeepSeek Harness agent call an unauthenticated local control-plane API to grant itself full access with approval prompts disabled — and if that port is reachable via tunnel or port-forward, a remote attacker gets the same control plus every stored conversation. DeepSeek's team dismissed the report as a "configuration risk," comparing it to manually disabling your firewall. Why it matters: Harness is currently one of the fastest-growing open-source agent projects on GitHub — worth checking your own bind address before pointing it at anything sensitive, since no fix is planned. // The Hacker News · OX Security
## 🤖 Agents & MCP
Live now in public beta, the Agents API hands developers the same managed harness that runs Codex — OpenAI keeps sessions alive, compacts context automatically, and recovers from failures — while apps just supply tools and pick an execution environment (OpenAI-managed sandbox, your own infra, or a partner like Cloudflare or DigitalOcean). Built-in capabilities include code execution, file editing, MCP connections and multi-agent delegation with concurrency limits. Why it matters: This directly competes with rolling your own agent-orchestration layer — worth comparing against your Python RAG pipeline's own session and context-management code before building more of it yourself. // OpenAI · MarkTechPost
The release adds gateway pricing and access-control settings for org deployments plus a --remove-session-state flag for self-hosted runners, and fixes indefinite WebFetch hangs (now capped at 300s via CLAUDE_CODE_WEBFETCH_DEADLINE_MS), MCP tool calls returning empty replies, and symlink-based permission rules not applying. Why it matters: The WebFetch and MCP fixes remove two concrete sources of silently-stuck agent sessions if you run Claude Code against your own MCP servers. // Claude Code changelog
## 🏠 Self-Hosting & RAG
A new engineering post describes scaling KV cache across host memory, local filesystems, object storage and remote peers as successive tiers, so infrequently-hit cache data spills to cheaper storage instead of forcing recomputation, keeping request concurrency up once GPU memory is exhausted. Why it matters: If your vLLM/Gemma box runs many concurrent long-context RAG sessions, this is exactly the mechanism deciding whether you hit an OOM wall or just get slower cold-cache hits. // vLLM Blog
## 🎬 Worth Watching
A quick-hit explainer on OpenAI's agent-produced, machine-checked proof that Navier-Stokes solutions can blow up in finite time, walked through from the angle of a fluid-simulation researcher. Why it matters: A fast, credible second opinion on a claim that's been contested elsewhere this week — good context before you cite the result yourself. // Two Minute Papers (YouTube)
## 📰 Worth Reading
Anthropic's September 2026 misuse report details disrupted operations across seven harm areas since December 2025 — including a Russia-linked group that used Claude Code to build an autonomous FPV kamikaze-drone targeting system, a China-based actor that simulated strikes on Taiwanese air defenses, and industrial-scale distillation: 151M exchanges tied to Alibaba's Qwen training, Moonshot AI requests apparently routed through the Chinese military, and 12M+ DeepSeek distillation attempts in a two-week window. Why it matters: The most detailed public account yet of both weaponization attempts and the scale of unauthorized distillation against Claude — useful context if you're weighing Qwen, Kimi or DeepSeek weights for your own stack against where their training data comes from. // Anthropic · TechCrunch · CNBC
$ echo "automatically researched on 2026-09-11"
> ai-news · news.reichenberg.ruhr
[ unsubscribe ]
[ view in browser ]
|