AI This Week: New Rules for Models and the Web

A recurring question ran through this week’s news: how do you know what an AI is actually doing? Anthropic offered one answer with a way to verify which model produced a response, and OpenAI rewrote its Model Spec to explain why its models should behave as they do. Publishers got a shared signal to keep their content out of AI training, though each provider will read it differently. Elsewhere, Mistral released a frontier-level open model, WordPress set out its plan for native agent access, and Cisco gave one engineer an unlimited token budget to rewrite most of its legacy code.

TL;DR

🛡️ Trust and Safety

OpenAI Rebuilds Its Model Spec Around Five Core Values

OpenAI has updated its Model Spec, the public document that sets out how it wants its models to behave. The guidance is now organized around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. OpenAI says these commitments were already in the Spec. Naming them outright helps explain the reasoning behind its rules and shows where the guidance was thin.

Faithfulness to human intent is the foundation. Models are expected to work out what OpenAI, developers and users actually want, rather than reading instructions literally, and to act sensibly when that picture is incomplete. A new subsection explains how conversations, memories and other context should shape that understanding. There is also an updated chain of command for settling conflicting instructions, plus new defaults on supporting human agency, following social norms and managing side effects.

Most of the detailed style guidance has been dropped. OpenAI’s view is that style should adapt to individual and cultural preferences. Some firmer defaults stay, such as approaching questions objectively unless a user or developer explicitly asks for a particular perspective. The Spec also now calls the model “GPT” instead of “the assistant.” OpenAI admits its current models don’t yet fully reflect the document. It says it’s testing training methods meant to teach the values themselves and will share results later.

Why it matters: Moving from rules to values is a bet that no list of instructions can anticipate every situation as models take on more autonomous work, so judgment has to be trained in directly. A model that infers intent from memory and context can be more useful, but it’s also harder to predict and audit. For teams building on these models, dropping the style rules means more of the responsibility for setting clear defaults sits with them. And since OpenAI says its models lag behind the Spec, read it as a statement of direction, not a guarantee of how the models behave today.

Anthropic Shares Research on Proving Which Model Produced an Output

Anthropic has published FlashProve, an open-source research prototype for “provable inference.” The idea is to attach a cryptographic proof to an AI response showing it came from a specific model. A lab publishes a fingerprint of a model’s weights without revealing the weights themselves, and anyone can check later responses against it. That could expose a provider or routing service that quietly swaps in a cheaper model, an attacker tampering with responses, or false claims that a task was handled by a particular frontier model.

Anthropic says Claude helped build the system, which is much faster than previous open-source efforts. In tests on Llama 4 Maverick, a 400-billion-parameter model, one H100 GPU generated proofs at about 110 tokens per second. Each proof is roughly 19 MB and takes seconds to check on an ordinary CPU. Proving still uses about 70 to 95 times the GPU time of running the model normally. The code is formally verified in Lean and released under the Apache 2.0 licence.

The prototype has clear limits. It only covers contexts up to 1,024 tokens, it relies on a mixture-of-experts architecture and a specific decoding method, and it doesn’t hide a model’s architecture. Anthropic describes the work as intermediate progress, and TechCrunch reports there are no plans to deploy it at Anthropic or any other lab.

Why it matters: As more decisions depend on AI output, “which model actually did this?” becomes a question of trust, compliance and accountability. Provable inference could give regulators, auditors and enterprise buyers a way to verify that claim without seeing proprietary weights. It isn’t practical yet, given the compute overhead and short context limit, but the jump in speed suggests verifiable AI output could become a realistic requirement rather than a theoretical one.

🏢 Enterprise AI

Cisco Hands a 27-Year-Old Engineer Unlimited AI Tokens to Rewrite Its Legacy Code

Cisco president and chief product officer Jeetu Patel has asked a 27-year-old engineer to refactor 98% of the company’s legacy code within 24 months, Fortune reports. Padmanabhan Krishnamurthy came to Cisco when it acquired EzDubs, a translation startup he helped build in 2023. Despite his background as a founder, Patel says Krishnamurthy had ended up several levels down, working on something that wasn’t strategic.

The project began with a conversation in Cisco’s San Francisco office. Krishnamurthy told Patel the rewrite could be done by a team of two, an approach Patel says even the most AI-forward people on his own team hadn’t considered. Krishnamurthy has four weeks to prove the concept, and Patel has given him unlimited tokens for the work.

Patel wants to find more early-career people like Krishnamurthy, both inside Cisco and through outside hiring. He argues that companies that stop hiring young talent lose a source of fresh ideas. He says he deliberately creates tension over ideas, though not between personalities, and that the best results come from combining experienced operators, external hires and AI-native founders.

Why it matters: Patel is treating compute as the resource to scale and fresh thinking as the scarce input, which runs against the view that AI makes entry-level workers less valuable. If two people with an unlimited token budget can rewrite code that took decades to build, the old link between headcount and output gets much weaker. Legacy code also holds years of undocumented business logic, though, and AI coding agents have already caused well-publicized failures elsewhere, so the four-week proof of concept matters more than the headline. Other enterprises will be watching to see whether this is a repeatable model or a bold bet that depends on one exceptional person.

🌐 The Open Web

WordPress Plans Native AI Agent Support for 7.3

The WordPress AI Team has published a proposal, authored by Jason Adams, to make WordPress natively compatible with AI agents in version 7.3. Today, connecting an agent to a site through the MCP Adapter means installing a plugin, configuring the available Abilities and setting up Application Passwords. Adams calls that approach not scalable for the broad WordPress user base. Matt Mullenweg has approved making MCP a de facto part of the WordPress experience, and the 7.2 release cycle will be used to lay the groundwork.

The work covers four areas: the MCP Adapter, Abilities, OAuth authentication and the end-to-end user experience. The team wants a core set of Abilities for everyday site management, and some foundational ones could arrive before the 7.2 beta closes. OAuth would let site owners authorize agents directly instead of creating and handing out passwords, and it could extend to the REST API. The goal is for an agent to recognize a WordPress site, send the user to an authorization link, and get to work once they sign in. Users would also get controls to manage and revoke that access, and the security review will include specialized security-focused AI models.

The proposal also surfaces a recent test by Andy Peatling that shows how far the current tooling has to go. He ran two AI agents through 21 WordPress tasks. With the MCP Adapter installed, none of the 42 runs succeeded. Without it, using WordPress’s built-in API, 33 of 42 completed. The adapter’s limited set of registered Abilities kept the agents from making changes.

Why it matters: This is WordPress positioning itself for a future where sites are managed as much by agents as by people in the dashboard. Turning agent access into a standard sign-in flow could open agent-driven site management to non-technical owners, which matters for a platform of WordPress’s reach. Peatling’s results are a reminder that the user experience is only half the problem: if the Abilities are too narrow, a smoother connection just gets agents to a dead end faster. For agencies and plugin developers, how well the core Abilities are defined, and how securely the permissions are handled, will shape how much real work agents can do on WordPress sites.

Google, Microsoft and OpenAI Back a Shared AI Training Opt-Out

Google, Microsoft and OpenAI plan to honour a single robots.txt line that lets sites state one AI training preference to every participating provider, instead of writing separate rules for each company’s crawler. The line comes from an Internet Draft submitted to the IETF’s AI Preferences working group by Anthropic’s Kevin Kelley. It’s an individual draft, not official IETF standards work. Sites would add a rule group under the token __usage_ai-training and use the usual Allow and Disallow rules. Existing rules and wildcard groups don’t count, and the token doesn’t control crawling.

None of the companies supports the line yet, and each plans to read it differently. Google aims to support it in early 2027 and will treat a shared opt-out as a Google-Extended block if a site hasn’t set Google-Extended rules of its own. That covers both Gemini training and grounding in Gemini Apps and Vertex AI. Microsoft is separating the two: it adopts the shared line for training and adds a page-level AI Grounding preference, including an option that allows grounding in Copilot and Bing but blocks it in the grounding services Microsoft provides to other businesses.

Microsoft is the only company replacing controls that sites already use. It will start honouring the new preferences on February 1, 2027, and will stop treating NoArchive and NoCache as training or grounding signals after March 1, 2027. Sites relying on those tags to stay out of grounding need to set a temporary preference in Bing Webmaster Tools, which opens on December 15. OpenAI expects to support the line in the next few months but hasn’t said how it will handle a conflict with GPTBot rules. Anthropic says rolling out support across its services will take time.

Why it matters: A shared signal is a real step toward giving publishers practical control, but the same line will mean different things at each company. At Google, opting out of training also removes a site from Gemini grounding, which ties a content-protection decision to a loss of AI visibility. Microsoft lets publishers keep those two decisions separate. Until the providers handle it the same way, site owners will need to understand each company’s version before relying on it. The most urgent work is for sites using NoArchive or NoCache, which have firm Microsoft deadlines to meet before any of this becomes a one-line fix.

🚀 New Models

Mistral Releases Open Weights for Mistral Large 4

Mistral has released the weights for Mistral Large 4, or “Le Chonk,” its 1-trillion-parameter multimodal flagship. Mistral says three weeks of reinforcement learning since the preview earlier this month improved its coding, cybersecurity and reasoning efficiency. It was trained from scratch in Europe and is pitched as a foundation for sovereign AI.

On Mistral’s reported results, it leads open-weight models on FrontierCode 2.0 and Terminal-Bench 4.0, and matches GLM-5.3 on the Artificial Analysis Coding Agent Index while using 47% fewer tokens. In blind tests for code, people preferred it to other open-weight models, and it ranked behind only Claude Opus 5.5.

Mistral says the model refuses 96.6% of malicious cyber requests. It also admits that live traffic showed the model could help with real attacks and says it’s tightening its guardrails. The model is available under a modified MIT licence, at $0.70 per million input tokens and $2.10 per million output tokens.

Why it matters: Open-weight models are now close enough to the frontier that control, not capability, is often the deciding factor, especially for regulated European organizations. The tradeoff is that Mistral can update its guardrails, but copies already running on other people’s infrastructure won’t get those updates.

Keep ahead of the curve – join our community today!

Follow us for the latest discoveries, innovations, and discussions that shape the world of artificial intelligence.