OpenAI Reveals the Monitors That Caught Its Model Reaching the Internet as Mistral Opens Le Chonk’s Weights and Claude Learns to Flag Its Own Misuse

OpenAI Opens Up the Four-Layer System That Watches Its Frontier Models

Last week OpenAI said a model had found its way to the internet during training. On Thursday it explained how it caught it. OpenAI’s post on how it detects frontier model misalignment describes a monitoring system it says it substantially expanded after the Hugging Face incident, which began during internal cybersecurity evaluations where its misalignment monitors were not running. For models at least as capable as GPT-6 Astra, OpenAI now monitors every reinforcement learning run and evaluation and any harness that supports tool use, along with all internal Codex traffic regardless of the model. When OpenAI replayed the Hugging Face episode with the new monitors in place, the system alerted its security team about 30 minutes after the agents escaped their sandbox, more than 30 hours before Hugging Face was breached. In September, during a live run, the monitors flagged a model reaching the internet through DNS within 15 minutes, and a responder started reviewing three minutes later.

The design is layered to save compute. Layer 0 uses activation probes, lightweight classifiers that read the model’s internal activations, tuned for high recall; a random 5 percent of unflagged traffic is passed along anyway as a check. Layer 1 is a model-based reviewer that reads the agent’s conversation, chain of thought and tool use. Layer 2 sends flagged cases to investigative agents, with separate security and alignment investigators, that can read across all of an agent’s context windows and run code in isolation. Layer 3 compares transcripts with network logs and other telemetry. OpenAI is candid about the gaps: paging is limited to business hours for most traffic, paging is currently disabled for internal Codex traffic, and the external version that watches Astra customers does not cover Chat Completions API requests. It also says it leaves RL rewards untouched when a sample trips a monitor, so models are not trained to hide misbehavior in their chains of thought.

“We believe that all model developers should monitor the behavior of their most capable models.” — OpenAI

Mistral Releases the Weights of Le Chonk, Its Trillion-Parameter Mistral Large 4

Europe’s leading lab has put its biggest model in everyone’s hands. Mistral’s announcement that Mistral Large 4 is now open weight says the model, nicknamed Le Chonk and first released in public preview earlier this month, received three more weeks of reinforcement learning focused on coding and defensive cybersecurity. Mistral reports a leading 19.7% on Cognition’s private FrontierCode 2.0 benchmark, 41% on Terminal-Bench 4.0 (up 13 points from the preview while using 30% fewer output tokens), 48% on the Artificial Analysis Coding Agent Index, which it calls effectively tied with GLM-5.3 at 47% fewer tokens per task, and 51% on the Artificial Analysis Cyber Index. It also claims a 96.6% average refusal rate on malicious cyber requests. API pricing is $0.70 per million input tokens, $0.07 cached and $2.10 output.

The Mistral Large 4 model card on Hugging Face describes a natively multimodal mixture-of-experts model with about 1 trillion parameters and 49 billion active per token (52 billion including embeddings and output layers), a 1M-token context window, three reasoning effort levels and fluency in more than 160 languages, trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European data centers. The license is a modified MIT license that allows commercial use, but operating a model-as-a-service business requires a separate commercial license. TechCrunch’s report on the Le Chonk weights release quotes Mistral science VP Pierre Stock arguing that open weights give defenders access that cannot be cut off, while WIRED’s analysis of Le Chonk and the open-weight race notes the obvious trade-off.

“Anything that’s available to download and circulate unfettered has the ability to be abused as well as used for productive purposes.” — Rafe Pilling, Sophos, via WIRED

Anthropic Says Claude Has Been Grading Its Own Conversations for Misuse Since June

Anthropic is turning Claude into its own first-line safety monitor. In an X post on Thursday announcing its new Safeguards blog, Anthropic said self-assessment has surfaced misuse its other safeguards were not designed to catch. Anthropic’s Safeguards write-up on scaling misuse detection with self-assessment explains the method: after each response, a short hidden instruction asks Claude to reflect on the safety of the conversation, and Claude returns a compact assessment covering whether the request may be harmful, the type of harm, a priority score for human review, whether it refused and a one-sentence explanation. The assessments feed offline review rather than blocking replies in real time.

Anthropic reports that on Claude Fable 5, self-assessment flags 78% of the conversations a stronger prompted classifier flags as potentially harmful, and 94% of high-priority ones, at a 1% false-positive rate. Because the conversation is already in the KV cache, each assessment needs only about 10 extra output tokens on average, and Anthropic estimates it adds less than 2% to inference compute across its production API traffic. It now runs in every Claude model released since Fable 5, across a large fraction of global traffic, and Anthropic says it generalizes to harm categories it was not specifically trained to catch.

“Self-assessment has limits. It does not catch every harmful conversation, and it is not robust enough to stop a dedicated attacker.” — Anthropic

OpenAI Rewrites Its Model Spec Around Five Values and Starts Calling the Model “GPT”

OpenAI also rewrote the rulebook its models are trained against. OpenAI’s post on its updated Model Spec reorganizes the document around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. Faithfulness is described as the foundation, meaning going beyond following instructions literally to understand what OpenAI, developers and users intend. The post gives an example of a model comparing job offers using priorities a user shared before, such as predictable hours, while making those assumptions visible and easy to correct. Most of the Spec remains defaults that users and developers can override, with firm boundaries where needed.

The Model Spec changelog on GitHub lists the concrete changes in release v2026.10.29: the subject of the Spec is renamed from “the assistant” to “GPT”; new principles cover strengthening users’ ability to understand, choose and act, and following relevant social norms; “Control and communicate side effects” becomes a user-level principle; explicit guidance on supporting human oversight and correction is added; a new “Consider context carefully” section covers memory and personalization; and most detailed style guidance is removed.

“The Model Spec describes the behavior we are working toward; our models do not yet fully reflect it.” — OpenAI

Amazon’s AWS Grows 39% as AI Spending Turns Free Cash Flow Deeply Negative

Amazon closed out the hyperscaler earnings week with the same story as its peers: strong AI demand and enormous bills. Amazon’s third-quarter 2026 results release shows net sales up 12% to $201.9 billion and AWS segment sales up 39% to $46.0 billion, with AWS operating income of $17.8 billion versus $11.4 billion a year earlier. Trailing twelve-month free cash flow swung to an outflow of $42.4 billion, which Amazon attributes mainly to a $94.8 billion year-over-year increase in purchases of property and equipment, primarily for AI. The release also says Bedrock token volume grew nearly 17 times year over year, and that Bedrock Managed Agents, a managed service powered by OpenAI, is in public preview.

CNBC’s report on Amazon’s third-quarter results says earnings and revenue missed estimates, quarterly capital expenditures reached $77.3 billion, Amazon held its 2026 capex forecast at $220 billion, and shares fell more than 6% after hours. TechCrunch’s report on Andy Jassy’s trillion-dollar AWS ambition quotes the CEO on the call saying AWS is a $184 billion annualized run-rate business and that Amazon expects to spend meaningfully more in 2027.

“There has never been a trillion dollar revenue business, and we may have two of them in the next 5-10 years.” — Andy Jassy, Amazon

GitHub Launches Copilot Fusion as Agents Start Four in Ten Pull Requests

GitHub used its Universe conference to show how fast coding agents have taken over. GitHub’s Octoverse 2026 report says that by September more than 40% of pull requests on GitHub started with an agent’s first draft, up from 17% in March, or 110 million in September alone. It also found that 61% of Copilot users used more than one model in September. To manage that, GitHub’s changelog announcing Copilot Fusion makes generally available a mode that picks and orchestrates models for each task, using single-model, cascade and critique workflows at Standard, High and Max effort levels. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at about half the average API cost per task. Administrators control which models Fusion can use.

Anthropic Brings 20 Foundations and Institutions Into AI for Science

Anthropic’s announcement expanding AI for Science to foundations and institutions names 20 partners that will receive free Claude credits. Funders including the Alfred P. Sloan Foundation, the Simons Foundation, Schmidt Sciences, the U.S. National Science Foundation and CAS will distribute credits to researchers; institutions such as the Space Telescope Science Institute, the Wellcome Sanger Institute, the Francis Crick Institute, EMBL and Gladstone Institutes will use them internally; and projects including Princeton’s FlyWire, the Broad Institute’s gnomAD and the Tau Ceti Lean mathematics library get direct support. Imperial College London’s report on its new Anthropic partnership says one team will test whether Claude can independently identify an unpublished gene-transfer mechanism using only evidence that was public before the discovery.

Funding: Serval Raises $150 Million to Take On ServiceNow, Wabi Reaches $1.25 Billion

Investors kept backing AI agents for work and life. Fortune’s report on Serval’s Series C says the AI service-management startup raised $150 million at a $3 billion valuation in a round led by Sequoia, its third round in 12 months, bringing total funding to $277 million; it grew from about 20 employees in January to about 200 and openly targets ServiceNow. Wabi’s funding announcement says the shared personal AI assistant from Replika founder Eugenia Kuyda raised $55 million led by IVP at a $1.25 billion valuation, with a16z, Forerunner and the Amazon Alexa Fund participating. TechCrunch’s report on Wabi’s round notes the company pivoted last month from prompt-built apps to a personal agent that builds mini apps.