OpenAI's Model Spec Update And NCSC's Agentic AI Guidance Send The Same Message To UK Businesses

Model Intelligence & News

23 August 2026 | By Ashley Marshall

Two of the most useful things that happened in AI this week had nothing to do with a new model launch. They were both about admitting limits.

OpenAI Just Told Its Models To Admit What They Cannot Do

On 18 August 2026, OpenAI published an update to its Model Spec, the living document that defines intended model behaviour. Buried among clarifications on how assistants should handle teen users and unsupported premises is a new section called "Be clear about capabilities and limits". It is a small addition in word count, but it is a significant admission: the company that makes the most widely deployed assistant in UK businesses is formally instructing its models to stop implying competence they do not have.

This matters because most UK firms using ChatGPT or the API have never had a clean answer to a basic procurement question: what does this thing confidently get wrong? Sales demos do not surface that. Marketing does not surface that. A published behavioural spec, updated in public with a changelog, is one of the few places a business can actually point to and say "this is the standard the vendor holds itself to".

The same update also removed outdated guidance written for pre-reasoning models, which is a reminder that behavioural expectations from even 18 months ago may no longer describe how the model you are using actually behaves. If your AI acceptable use policy or vendor risk assessment cites model behaviour documentation from last year, it is worth checking whether the underlying spec has moved since you last read it.

The practical takeaway is not that OpenAI has solved honesty. It is that there is now a public, dated, versioned commitment you can hold a supplier to, and a changelog you can monitor. Few UK firms currently do either.

NCSC's Warning: Scope The Autonomy Before You Grant It

Two days after OpenAI's update, the National Cyber Security Centre published a blog titled Managing the cyber risk of agentic AI, describing it as interim advice while formal guidance is developed with industry partners. Coverage in Infosecurity Magazine summarised the core recommendation plainly: firms should first assess how much autonomy a system actually needs, and identify what could go wrong, before deployment.

The NCSC's framing draws a distinction that many UK businesses currently collapse into one category called "AI agent". An agent that suggests options to a human carries a fundamentally different risk profile from one that touches production systems and decides without a person present. The NCSC's position, reported by industry outlets covering the interim advice, is that controls should scale to that gap rather than being applied uniformly across every deployment labelled "agentic".

This is the security equivalent of OpenAI's capability honesty principle. Where OpenAI is asking the model to be honest about what it can do, NCSC is asking the business to be honest about what it is actually letting the model do. Most incident write-ups from agentic AI failures in the past year share a common root cause: the autonomy the system was given in practice exceeded the autonomy anyone had consciously decided to grant it. It crept in through default settings, permissive API scopes, or a "quick pilot" that was never formally reviewed before it touched a live customer record or a live payment.

NCSC's advice is interim precisely because this is moving fast. That is not a reason to wait for the formal version before acting.

Twelve New Models In One Month Make The Honesty Problem Harder, Not Easier

Both of these announcements land in the middle of an unusually fast release cycle. According to LLM Gateway's model release timeline, 12 new AI models shipped in August 2026 alone, from 7 different providers, as of 17 August. That includes GLM-5.2 Turbo and GLM-5.3 from Z.AI, Gemini 3.7 Flash from Google, Grok 4.6 from xAI, Seedance 2.5 and Seed 2.1 Turbo from ByteDance, Muse Spark 1.2 from Meta, and Qwen3.8 Max from Alibaba, described elsewhere as one of the largest open-weight releases to date. OpenAI itself is mid-rollout of GPT-5.6 Sol to eligible paid ChatGPT plans, while simultaneously retiring OpenAI o3 from ChatGPT on 26 August 2026 after a 90-day sunset period.

Every one of those releases arrives with its own capability claims, its own benchmark chart, and its own gaps that will not be fully understood until businesses have used it in anger for a few months. A UK firm evaluating models this quarter is choosing between more genuinely capable options than at any point before, but also inheriting more unverified capability claims than at any point before.

This is exactly the environment where capability honesty from vendors and autonomy discipline from buyers stop being nice-to-haves and start being the only practical defence. You cannot manually red-team 12 new models a month. What you can do is insist that any model your business adopts publishes a behavioural spec you can check, and that your own deployment never grants more autonomy than you have deliberately reviewed, regardless of how good the benchmark chart looks.

What Capability Honesty Actually Looks Like In A UK Customer-Facing Assistant

In practice, capability honesty is not a philosophical stance, it is a small number of concrete behaviours you can test for. Does the assistant say "I don't have access to your order history" rather than guessing plausible order details? Does it decline to give financial, legal, or medical specifics it is not qualified or authorised to give, rather than answering confidently and wrongly? Does it distinguish between "I know this" and "this is my best estimate" in its phrasing, especially for numbers, dates, and policy details a UK customer might rely on?

Most UK businesses using AI customer service tools have never actually tested for this. They test for tone, for resolution rate, for whether the bot escalates angry customers correctly. They rarely test the failure mode where the assistant is calm, fluent, and confidently wrong about a return policy, a delivery date, or a regulatory position, because that failure looks identical to a correct answer until a customer acts on it.

This is where OpenAI's Model Spec change becomes operationally useful rather than just a public relations note. It gives you language to put in a supplier questionnaire: does the model behaviour spec explicitly require capability honesty, and can you show us the version and date of the spec you are running against? A vendor that cannot answer that in writing is one you are trusting on tone of voice, not on tested behaviour.

For UK firms in regulated or reputation-sensitive sectors, this single test, run against your live production prompts on a rolling basis, catches more real customer-facing risk than most existing QA processes.

Building An Honesty And Autonomy Check Into Your Next AI Rollout

Put the two threads together and a practical, lightweight check emerges for any UK business about to expand its use of an AI assistant or agent. Before the next rollout, expansion, or model swap, answer four questions in writing and keep the answers somewhere a colleague or auditor could actually find them.

First, what does this system explicitly say it cannot do, and where is that documented by the vendor, not assumed by your team. Second, what is the actual autonomy this deployment has today, not what you intended when you configured it, including any default permissions or API scopes that may have widened since launch. Third, what specifically could go wrong if this system's confidence and its actual competence diverge in a live customer or financial interaction. Fourth, who reviews and signs off any change to either the model version or the autonomy scope, and how often does that review actually happen versus how often it is supposed to happen.

None of this requires a large governance programme. It requires a short document, updated whenever you change model or expand what the system is allowed to touch, and a habit of checking vendor behaviour specs and NCSC advisories with the same regularity you check a supplier's SLA. Given that NCSC has explicitly said the current advice is interim while formal guidance is developed, firms that build this habit now will find the eventual formal guidance easy to slot into an existing process rather than a fresh compliance exercise.

The Pushback: "Our Vendor Already Covers This In The Terms"

The most common objection to this kind of check is that it duplicates what is already in a vendor's terms of service or acceptable use policy. It generally does not. Terms of service protect the vendor from liability. A behavioural spec like OpenAI's, and an autonomy scoping exercise like the one NCSC recommends, are about protecting your customers and your business from a failure mode occurring at all, not about establishing whose fault it was afterwards.

A second pushback is that this slows down adoption at exactly the moment competitors are moving fast on 12 new models a month. The counter to that is straightforward: the firms most likely to have a public, damaging AI failure in the next 12 months are the ones treating every new model as a drop-in upgrade with no fresh autonomy or honesty check, precisely because the pace of releases is now too fast for informal, ad hoc trust to keep up. A five-question review before go-live is not a brake on adoption speed, it is what makes fast adoption survivable.

The businesses that will get genuine value from this month's model wave, rather than a headline about it going wrong, are the ones treating vendor capability claims and their own deployment scope as things to be checked on a schedule, not assumed at launch and forgotten.

Frequently Asked Questions

What exactly did OpenAI change in its Model Spec on 18 August 2026?

It added a new section titled "Be clear about capabilities and limits", clarified how assistants should handle false or unsupported premises, updated guidance on appropriate relational interactions for teen users, and removed outdated instructions written for pre-reasoning models.

Is NCSC's agentic AI guidance final or still being developed?

It is explicitly interim. NCSC has said it is working with partners to develop formal guidance that will build on and eventually supersede the current blog-form advice, so UK businesses should expect the detail to evolve.

What is the core recommendation in NCSC's interim advice?

Assess how much autonomy an AI system actually needs for its task, and identify what could go wrong, before deployment. Controls should scale with how much a system can decide or act on its own, not be applied uniformly to every system labelled agentic.

Why does capability honesty matter more with 12 new models released in a month?

Each new model arrives with unverified capability claims that take months of real use to properly understand. When adoption speed outpaces verification speed, a vendor's own honesty about limits and your own scoping of autonomy become the only practical safeguards available in the short term.

How can a UK business actually test an assistant for capability honesty?

Run live production prompts through it and check whether it distinguishes clearly between things it knows, things it is estimating, and things it does not have access to, particularly around order data, policy details, dates, and figures a customer might act on.

Does this apply to businesses only using consumer ChatGPT, not building custom agents?

Yes. Capability honesty in the underlying model behaviour affects any customer-facing use, including simple chat deployments, not just custom agentic workflows with tool access.

What is OpenAI retiring alongside these changes?

OpenAI o3 is being retired from ChatGPT on 26 August 2026 following a 90-day sunset period. GPT-4.5 was already retired from ChatGPT on 27 June 2026. Both changes apply to ChatGPT only, with no changes to API access.

Where should a UK business look for updates to both threads?

OpenAI publishes Model Spec changes in its Help Center release notes, and NCSC publishes agentic AI advisories on its blog. Both are worth adding to whatever supplier or regulatory monitoring list your business already keeps.