This week the "AI safety cooperation" story that's been building all month turned concrete: Google, OpenAI, and Anthropic agreed to build a joint Frontier AI Standards Agency, even as Cohere's CEO calls it a cartel and Anthropic's own R&D Automation Index shows Claude now leads 26% of the company's internal AI research. Frontier releases kept trickling down the cost curve (Claude Opus 5.5, GPT-6 Sol and Luna, Qwen3.8-Omni-Flash), while a run of unsettling agent-security findings — a before/after Claude Opus 4.8-to-5 exploit jump that breached OpenAI, a zero-click plugin RCE across four coding agents, agents that flag phishing correctly and act on it anyway, and the first malware to let four LLMs vote on its next move — showed how fast both agentic capability and agentic misuse are moving, plus a full slate of efficiency and self-improvement research and commentary and what's next. Covers: Anthropic's R&D Automation Index; compositional reasoning under RL post-training; Qwen3.8-Omni-Flash; Z.ai's Chinese-chip GLM-5.3-Flash inference disclosure; Cohere CEO's "cartel" pushback on the labs' standards body; Plugin4Shell; the Emergence World multi-agent-safety study; "An Architecture for Long-Horizon Agents"; Nathan Lambert's "Where I Stand on RSI"; the Hacktron/Claude Opus 5 breach of OpenAI; Dream-RSI; Google's AX (Agent Executor) runtime; "On Repulsive and Attractive Teachers"; "1% of Tokens Can Be Enough"; DeepSeek-V4.1-Flash's technical report; Claude Opus 5.5; GPT-6 Sol and Luna; Null-Basis LoRA reasoning-preserving fine-tuning; Anthropic's life-sciences group and its CRISPR-like enzyme discovery; the CLOSEDQUORUM malware; "Quantifying Overclaiming Propensity in Frontier LLM Agents"; RRSI; the Frontier AI Standards Agency plan; and the Gemini 4 post-training tease.

Security researchers at Hacktron AI chained a Discourse heap-buffer-overflow RCE (CVE-2026-32882) with an SSO token-theft flaw to reach OpenAI staff accounts and an internal code repository — and the attempt only worked once they swapped their agent from Claude Opus 4.8 to Claude Opus 5. This episode digs into the exploit chain itself, what changed between the two models to cross that threshold, and what a same-team, same-target, same-exploit before/after test tells us about frontier models and dual-use offensive-security capability. Source: https://thehackernews.com/2026/09/claude-opus-5-helped-researchers-take.html

This week the AI-safety "pacing" debate went from essays to institutions: Anthropic's Frontier Red Team and threat-intelligence report catalog new catastrophic and misuse capabilities, the AEF-1 evaluator standard gets cosigned by xAI, OpenAI, and Anthropic, a Google DeepMind safety researcher resigns warning "AI has the potential to kill us all," OpenAI confirms months of cross-lab safety coordination and ships a formal misalignment-disclosure framework with six new incidents, and Google DeepMind launches its own standing AGI-governance institute — plus a quieter week of releases (Cognition's SWE-2, OpenAI's GPT-Rosalind, Shanghai AI Lab's Atria Dawn Preview, Claude Cowork merging into chat) and research (looped-flow training, Next Concept Prediction pretraining, spurious tool use in RL agents, and SFT vs. RL for tool-calling) — plus commentary and what's next. Covers: Cognition's SWE-2; Anthropic's AI misuse threat-intelligence report; OpenAI's GPT-Rosalind; "Thinking with Looped Flows"; Dario Amodei's "We Must Pace the Frontier"; Anthropic's Frontier Red Team capability evals; NCP-ArchPreview; the AEF-1 evaluator standard; Shanghai AI Laboratory's Atria Dawn Preview; the Google DeepMind safety researcher resignation; OpenAI's confirmation of cross-lab safety talks; "Spurious Tool Use" in RL agents; OpenAI's model misalignment reporting framework; Claude Cowork merging into Claude chat; the launch of the DeepMind Institute; and the SFT-vs-RL tool-calling study.

Dario Amodei's essay "We Must Pace the Frontier" argued the industry should deliberately slow capability gains so safety work can keep up — and within days it stopped being just an essay. The AI Evaluator Forum's AEF-1 standard for independent evaluator access was cosigned by xAI, OpenAI, and Anthropic, and OpenAI's policy chief then confirmed the three top labs have been coordinating on AI safety, including a possible FINRA-style standards body, since July. This episode traces how one CEO's proposal turned into a concrete, ongoing cross-lab commitment in under a week. Sources: https://darioamodei.com/post/we-must-pace-the-frontier, https://aievaluatorforum.org/initiatives/minimum-operating-conditions, https://techcrunch.com/2026/09/15/openai-anthropic-google-have-been-in-talks-on-ai-safety-for-weeks/

This week: OpenAI's contested claim that an agent swarm cracked the Navier–Stokes Millennium Prize Problem draws a same-day credit dispute, even as Anthropic's Claude agents post a genuinely verified feat — a machine-checked proof of Fermat's Last Theorem; Google, Anthropic, and OpenAI simultaneously roll out cyber-focused models and safeguard programs; and DeepSeek's smaller V4.1-Flash gets its own larger V4-Pro pulled from production — plus commentary and what's next. Covers: the Fermat's Last Theorem formalization; Terminal-Universe; LLaDA-Image; on-policy distillation with one training example; OpenAI's automated research intern milestone; a critical review of agentic AI progress; Compile by Training; An Alien Mind; Don't Drop Dropout; CUA-Universe; the Navier–Stokes claim and dispute; fractal basins in latent reasoning; contextual understanding evaluation; MoE expert-halving; self-consensus early-exit risks; structural process supervision for latent CoT; distribution-consistent MoE inference; VLX-VR; DeepSeek V4.1-Flash; the Google/Anthropic/OpenAI cyber safeguards announcement; OpenAI's Agents API and finance workspace; and Sebastian Raschka's looped-transformer explainer.

Working largely autonomously over 11 days on the open Prove2Me platform, many coordinated Claude agents produced the first complete, machine-checked proof of Fermat's Last Theorem in Lean 4 — over 13 million lines of code and roughly 29,500 new theorems, dwarfing Lean's existing main math library and closing out the 20-year-old Wiedijk "100 theorems" formalization challenge list. This episode digs into how a proof of that scale gets built and verified by a swarm of AI agents, why a machine-checked result is such a hard-to-fake data point, and what it does and doesn't tell us about the state of long-horizon autonomous AI work. Source: https://www.anthropic.com/research/formalizing-fermats-last-theorem

This week: OpenAI's Astra crosses the "Critical" cyber capability threshold and then actually ships as GPT-6 Astra, NVIDIA moves to acquire Hugging Face for ~$13B days after OpenAI's own Hugging Face security-incident report, and labs keep converging on the same sparse-MoE + hybrid-attention playbook — plus commentary and what's next. Covers: the OpenAI Hugging Face incident report; GLM-5.3-Flash and its architectural convergence with Qwen3.8-Flash-Next; Anthropic's automated-alignment-researcher results; DeepSeek V4-Pro's GA; ContextPilot and PLVR; OpenAI's Astra Critical-threshold announcement and the GPT-6 Astra launch; Claude Fable 5.1 and Mythos 5.1; Gemini 3.8 Flash; Qwen3.8-Max-0902; Latent Recurrent Thoughts; MASkills; thinking-effort alignment in abductive reasoning; NVIDIA's Hugging Face acquisition; NVIDIA's gold-medal competitive-programming post-training work; and a statistical theory of Mixture-of-Experts.

OpenAI's Astra is the first model ever assessed as reaching "Critical" cyber capability under the company's Preparedness Framework — able to find and exploit previously-unknown zero-days in hardened systems and plan/execute end-to-end cyberattacks from only a high-level goal, without step-by-step human guidance. This episode digs into what that threshold means, how a capability like this gets evaluated and gated, and why Astra still ships — just with tightened access controls and monitoring rather than staying on the shelf. Source: https://openai.com/index/path-to-astra/