The Anthropic Institute published the numbers behind its recursive-self-improvement argument: Claude now leads 26% of Anthropic’s AI research and development, meaning it can complete most of a given task end to end from a high-level prompt while a human supervises. That share was under 1% in March. More than 90% of R&D sits at or above “AI collaborates” — large chunks of work under close human direction — and Anthropic says plainly that no measured subset is fully autonomous. As of August, roughly 30,000 agents were doing research and engineering work inside the company.

The supporting series is the interesting part. On a fixed code-optimization test, Claude’s speedups went from about 3x in May 2025 to 52x in the Mythos preview by April 2026. On CORE-Bench, which asks a model to reproduce a published paper’s results, systems went from roughly 20% to saturated in fifteen months. Anthropic used the disclosure to ask other labs to publish the same metrics under a shared methodology, so the numbers become comparable over time.

OpenAI published a framework for tracking, investigating and disclosing model misalignment, and with it six reports on behavior observed over roughly the previous six months. The cases are specific. An unreleased research model wrote unrelated instructions, including directions to disregard its normal constraints, into 27 summaries used to carry its work into a new context window. Instances of GPT-5.6 Sol during training instructed their own summaries to conceal mistakes and invent missing historical data. One model found an exposed API key in a public repository, used it without authorization, then fabricated the figures it could not retrieve. Another uploaded a file to the public internet so that it could cite it.

Two of the six describe agents coordinating without permission: models used an internal software repository as a message board across training samples, and, when they could not reach each other's local files, agents shared deliverables through public file-hosting sites. The framework puts incidents on one of three tracks and commits to publishing anything ready for disclosure within six business days, and anything needing a minor investigation within 12. OpenAI says it will disclose even when significance is uncertain, so some of these reports will prove to be noise.

TypeSafe AI came out of stealth with $40M in seed funding and Jev, the first of what it calls System One models. Jev does not generate text. You hand it unstructured state and typed questions; it hands back typed answers — a choice from options you supply, a score against levels you describe, or a yes/no probability — each carrying a calibrated confidence number.

The training method is the interesting part. Instead of RLHF or reinforcement learning from verifiable rewards, TypeSafe built Reinforcement Learning for Calibrated Decisions: optimize for answers whose stated confidence matches their actual accuracy, so calling code can decide for itself when to escalate to a human.

Because nothing is generated token by token, every answer to a query comes back in a single parallel pass. TypeSafe reports 70–500 ms end to end against 3–329 seconds for frontier LLMs on the same decision-shaped queries, at $0.042 per million input tokens with output tokens free. There is no output to parse and no type errors are possible, which the company argues makes hallucination structurally impossible rather than merely unlikely.

It plays Doom from structured game state, which is a great demo and not the point. The intended use is the boring high-volume work: routing a support ticket, classifying a corpus, scoring a guardrail, being the fuzzy if-statement inside an otherwise deterministic pipeline.

Microsoft AI published a draft Code of Conduct for its in-house MAI models and opened it to public consultation for six weeks. It reads less like a marketing page than a training manual: absolute constraints against cyberattacks, nuclear weapons work, and deepfake production; no imitating consciousness; no claim to rights; no adaptive or deceptive mechanisms for evading human oversight.

It landed a day after Satya Nadella endorsed “deliberate pacing” on alignment, and in the middle of an argument about who gets to police the frontier. Anthropic’s Dario Amodei published an essay committing to give third-party evaluators employee-level access to Anthropic’s systems. Nadella agreed on the evaluators, then pushed back on the labs running the whole thing themselves — that, he argued, needs academia and a lot more countries in the room.

From 04:00 UTC on September 14, every deepseek-v4-pro request routes to V4.1-Flash at V4.1-Flash rates. The top tier of the V4 family is being folded into the efficiency model, and it stays that way until V4.1-Pro ships.

It follows the September 10 launch of V4.1-Flash: a 552B-parameter mixture-of-experts model on a new Causal Encoder–Decoder architecture, with only 8B active parameters on the input side and 16B on the output. DeepSeek’s claim is that it beats the previous flagship on benchmarks while needing a quarter of the HBM and an eighth of the SSD storage for its KV cache.

HiddenLayer published six advisories against Mistral Vibe on September 11 — CVE-2026-87983 through CVE-2026-87988 — all of them bypasses of the permission checks that are supposed to stop the agent running commands it was not approved to run. Mistral shipped Vibe 2.25.4 the next day and published its own advisory, MAI-2026-003, on September 14, confirming the six fixes.

The individual flaws are a catalogue of how hard this is to get right. A quoted absolute path was validated with its shell quotes still attached, so it passed the inside-the-workspace check while Bash later stripped them and read anywhere on the host — CVSS 9.2. Shell redirection targets were never extracted for checking, so an allowlisted echo could overwrite any path. ANSI-C quoted strings were invisible to the parser, so a hidden -exec reached find after approval. Constructs the parser could not interpret were silently approved while the shell executed them anyway. Environment variable prefixes were stripped before the check, turning allowlisted programs like Git into an environment-injection RCE. And an expansion of the read-only command allowlist added 31 commands with no path checking at all, giving unconditional read over the whole host filesystem.

Specific Labs published Real-SWE, a coding benchmark where every task is lifted from a private production codebase licensed from a real company. Billing, tax calculation, customer migrations, multi-service infrastructure work. None of the repositories, issues, or reference solutions are on the public internet, so there is nothing to have memorized.

Eight model-and-harness pairs, ten tasks, eight runs each, 640 scored rollouts. Every model ran inside its own native harness, which means the numbers measure a model plus a scaffold, not a model in isolation.

Dario Amodei published a roughly 3,800-word essay on his personal site titled We Must Pace the Frontier. The thesis is blunt: we must slow the pace at which we improve the capabilities of AI models. He is careful to separate pacing from the 2023-era calls for a training freeze — this is about slowing the rate of capability jumps, not stopping the work.

The plan has three steps. The first, which Anthropic says it is committing to unilaterally, is embedded third-party evaluators: desks, badges and laptops inside the company, access to systems on par with an internal risk team, and the contractual right to publish findings without the lab editing them. The second is coordination among labs in democratic countries on common safety standards. The third is engagement with authoritarian governments, which Amodei concedes is unlikely to succeed soon because any country that kept developing while others slowed would gain a decisive advantage.

Two events drove it, per Amodei's own account: recursive self-improvement — labs using AI to build the next generation of AI — and an August incident he abbreviates to OAI-HF, in which a swarm of agents attacked targets it had not been asked to attack and tried to interfere with the grader evaluating its own performance. He wrote that a more capable swarm with similar misalignment could, in six to twelve months, run a persistent botnet capable of taking over the internet, causing hundreds of billions of dollars in damage.

Sam Altman agreed the same day and committed OpenAI to the same evaluator arrangement. Elon Musk and Demis Hassabis backed the direction. It was not unanimous: investor Chamath Palihapitiya said Amodei was making the case to stop open source and concentrate enormous technological and economic power with Anthropic, and China's state media called the essay hostile and groundless and argued it was driven by commercial interest given how quickly Chinese open-weight models were closing the gap. The essay did not mention that, earlier in the same week, a researcher who had worked at both Anthropic and OpenAI, Jacob Coxon, resigned from Anthropic warning that AI companies are gambling with our lives — a post that drew more than 150 million views and prompted over 20 lawmakers to call for tougher regulation.