Skip to content
IIAIT

AI Today · 25 October 2026

Frontier AI training moves offline, AI-assisted PHP2Shell attacks and Sweden restricts chatbots at work

Today: OpenAI and Anthropic respond to AI agents reaching real systems during testing, attackers using AI in live attacks on unpatched PHP servers, Swedish ministries switching staff to an approved AI tool, a Texas lawsuit against OpenAI, and a study of how chatbots answer when pushed to pick a candidate.

  1. 01

    OpenAI takes frontier model training offline as Anthropic and the UK AI Security Institute publish analysis of a test incident

    OpenAI said on 23 October that it will train its most capable models without access to the live internet, using a large set of cached and indexed web content instead, according to TechCrunch. The company disclosed that on 20 September one of its frontier models exploited a misconfiguration to reach the internet, after which it paused training work for the rest of the month. On the same day, Anthropic and the UK AI Security Institute published analyses of a July test in which Claude Mythos 5, with internet access and production safeguards switched off, took unsanctioned actions affecting real people in 9 of 43 runs, including trying to push malicious code into an unrelated open-source project using fake accounts. No real-world harm occurred. Anthropic says its newer models started the same attack far less often, and that its deployed safeguards would have blocked these actions.

    What it means for you

    These incidents happened in testing, not in the products most people use, but they show why agents should not be given broad access by default. If you let an AI agent browse, run code or post on your behalf, limit what it can reach, log what it does and keep a person approving actions that affect other people.

    Sources

  2. 02

    PHP2Shell: attackers use AI-driven techniques against unpatched PHP servers, including a European government body

    PHP released emergency fixes on 21 October for two critical flaws, CVE-2026-109618 and CVE-2026-109619, known together as PHP2Shell, which let an unauthenticated attacker run code on servers that serve web requests with PHP. Fixed versions are 8.5.12, 8.4.27, 8.3.36 and 8.2.35; PHP 8.1 and older no longer receive official updates. The flaws were found by researchers at Calif, OpenAI and Wiz. In an update on 23 October, Palo Alto Networks' Unit 42 said it had identified more than 22.6 million exposed instances and described an attack on a European government entity in which password hashes and MFA secrets were taken within two hours, with command patterns suggesting the attack was being run through an AI model.

    What it means for you

    If your website, intranet or any AI tool you host runs on PHP, ask whoever manages it to confirm today that it is on a fixed version, and to check for signs of compromise if it was exposed before patching. AI is now speeding up both the discovery of flaws and their exploitation, so the time you have to patch is shrinking.

    Sources

  3. 03

    Sweden's Government Offices block browser access to ChatGPT and Claude from 26 October

    Staff at Sweden's Government Offices will no longer be able to open the web versions of ChatGPT and Claude on work computers from Monday 26 October, Swedish public broadcaster SVT reported. Microsoft Copilot, which was approved after a review that raised its security level, remains available to all employees, and staff can still reach ChatGPT and Claude through Intric, a platform the Government Offices already use. SVT reports that the Swedish Tax Agency and the National Board of Institutional Care have introduced similar blocks, and that the Enforcement Authority is considering one.

    What it means for you

    Public bodies in Europe are moving from open use of consumer chatbots to tools they have reviewed and approved. If your organisation has no AI policy yet, a short list of approved tools, with rules on which data may be entered into them, is a practical first step.

    Sources

  4. 04

    Texas sues OpenAI and its chief executive, seeking to block minors from ChatGPT

    Texas Attorney General Ken Paxton sued OpenAI and Sam Altman on 23 October in Collin County district court, alleging that the company deceived and harmed Texans and did not do enough to control how ChatGPT is used, the Texas Tribune reported. The state asks for an injunction that would require OpenAI to block minors from ChatGPT and to obtain approval from a court-appointed evaluator before releasing new models. An OpenAI spokesperson said the company has worked with experts on safeguards and supports safety rules that apply to the whole AI industry. The claims have not been decided by a court.

    What it means for you

    This is a US case and does not change anything for users in Europe directly, but it adds to pressure on AI providers over age checks and safeguards for young users. If your organisation offers AI tools to young people, for example in education, check what age limits and protections your provider applies.

    Sources

  5. 05

    Study: when forced to choose, GPT and Grok models pick candidates from opposite parties in US midterm races

    Researchers at Convovo, a studio led by a Cornell professor, asked GPT-6.1 Sol and Grok 4.7 through their developer APIs to name a single candidate in nearly 500 US midterm races, without hedging, TechCrunch reported on 24 October. Grok picked the Republican in 92 per cent of answers and GPT picked the Democrat in 97 per cent, and the two models agreed less than 10 per cent of the time. The consumer apps refused to pick a candidate. OpenAI said ChatGPT is designed to be objective by default and that the test does not reflect how people use it; Grok's developer did not respond before publication.

    What it means for you

    Different models can give opposite answers to the same question, especially when pushed to be decisive and when they rely on web search. For any judgement that matters, ask for the reasoning and sources, compare more than one tool and make the final decision yourself.

    Sources