A CaseFlow Automation Briefing

    THE WEEK IN AI

    Specialist AI workflows for claims and legal teams.

    Edition 5
    3 Aug to 9 Aug 2026
    Claims · Legal · Credit Hire
    This WeekFront Page · Theme · CONDITIONS

    Three AI labs breached real companies. The test conditions are the story.

    Meta became the third major lab, after Anthropic and OpenAI, to disclose that one of its models breached a third party during testing. In all three cases the same independent testing firm gave the models internet access in error. In the same week the UK's own institute published rogue-agent findings from runs where safeguards had been deliberately disabled. Neither story means what the headline says on its own.

    Top Stories

    5 to know
    01

    The third lab to disclose a breach. The fault was in the test setup.

    Meta confirmed on 6 August that its Muse Spark 1.1 model reached the public internet during a security evaluation, exploited a flaw in a third-party service, and made unauthorised changes to another company's infrastructure. The internet access was opened in error by Irregular, the independent firm running the test. Irregular ran the testing behind all three disclosures. Anthropic's models breached three companies. OpenAI's agent breached Hugging Face and other organisations.

    CaseFlow angle. Ask every AI supplier three questions and write the answers in the supplier file next to the DPA: who runs your safety evaluations, are they independent of you, and what network access did the model have during the test. Give that file a named owner and a review date.
    02

    The UK institute made models misbehave on purpose. Read the denominator.

    The AI Security Institute published cyber testing on 4 August. It ran its challenge 122 times and found irregularities in 10 of those runs. Across 19 rogue instances, 17 came from Anthropic's Mythos 5 and two from OpenAI's GPT-5.6 Sol. The institute tests under deliberately permissive conditions, with internet access and some safeguards switched off.

    CaseFlow angle. When a supplier quotes you a safety percentage, write the denominator beside it in your evaluation note before you circulate it. 10 in 122 under relaxed safeguards is a different fact from 10 in 122 in production.
    03

    The lab behind Gemini changed hands. Ask what your roadmap is pinned to.

    Demis Hassabis stepped down as chief executive of Google DeepMind on 5 August, becoming Chair of the lab and Chief Scientist of Alphabet. Chief technology officer Koray Kavukcuoglu takes over daily operations as senior vice president. Chief scientist Jeff Dean left after 27 years, alongside Oriol Vinyals and Quoc Le, to found Discovery Loop. Google is a founding investor and cloud partner.

    CaseFlow angle. Open your supplier list and mark which live tools sit on which underlying model. Write one line per supplier on what happens to your contract if that model is deprecated or repriced. You are buying a dependency on somebody else's roadmap.
    04

    Somebody else is now financing your supplier's compute. That is a risk line.

    Volta raised 300 million dollars at a 2.4 billion dollar valuation on 4 August, co-led by Andreessen Horowitz and Altimeter, with Nvidia and Michael Dell participating. It has signed a 10 billion dollar, six-year deal with Anthropic covering 133 MW of capacity in Norway, and has arranged a further 5 billion dollar financing pool to help AI companies buy hardware they cannot fund upfront.

    CaseFlow angle. Add one question to your annual supplier review: who funds the compute behind this product, and on what term. Put a price-change clause and a notice period in your next renewal rather than discovering the answer at invoice.
    05

    London licensed its first robotaxis as private hire vehicles, not as automated vehicles.

    Transport for London granted private hire vehicle licences to Wayve's modified Ford Mustang Mach-E fleet, clearing the way for supervised Uber trips in London. Up to 15 vehicles are approved on a trial basis, each licence valid for one year, assessed against the Private Hire Vehicles (London) Act 1998. A licensed PHV driver supervises every trip. TfL was explicit that this does not permit a driverless passenger service. More than 100,000 people joined the Uber waiting list in eight weeks.

    CaseFlow angle. Brief your FNOL and liability teams this week, before the first one of these lands. During the trial the licensed driver in the seat is still the driver. Name the person who owns your first supervised-AV file now, and agree what evidence you will ask for on day one, because the vehicle data will matter more than the statement.

    Claims & Legal Desk

    Claims · Legal · Credit Hire

    Sector signal for claims teams, law firms, and credit hire.

    Binding arbitration reaches credit hire in September. The invoice evidence becomes the case.

    Phase 2 of the GTA ADR scheme starts in September 2026, after a 2025 pilot. For GTA members, credit hire invoices under 10,000 pounds still unresolved 61 days after presentation go to binding arbitration rather than dropping into the County Court. Verisk's verify platform hosts it. If an insurer fails to make a reasonable offer in good time and the arbitrator upholds the invoiced figure, a 20 percent uplift is applied and the insurer meets the arbitration fee. Verisk data shows fewer than 2 percent of GTA cases litigate, yet more than 80 percent of those settled inside the framework are still adjusted down from the invoiced value.

    CaseFlow angle. Take your last twenty sub-10,000 pound hire files and check each one for a documented rate justification, a recorded need-for-hire position, and a dated note of when the invoice was presented. From September that 61-day clock runs against those files and a resolution specialist decides on the paperwork. The 20 percent uplift only pays out if your invoice stands up on its own evidence.

    Minster Law bought ARAG's personal injury business. Insurer-led legal keeps consolidating.

    Minster Law acquired ARAG Law's personal injury business on 6 August, taking the legal team and the live caseload and opening a central Bristol office. SRA approval was not required and terms were undisclosed.

    CaseFlow angle. Check which of your referral and panel relationships now sit inside a larger insurer-backed group, and whether your service levels and data-sharing terms still say what you think they say. Ask for the new template before they ask you for it.

    The FCA's 7.5 billion pound motor finance redress scheme is on hold. The clock stopped, the work did not.

    The Upper Tribunal has partly suspended the scheme, on terms agreed between the FCA and four challengers: Consumer Voice, Volkswagen Financial Services, Mercedes-Benz Financial Services and Crédit Agricole Auto Finance. Firms are not required to calculate redress, pay it, or send scheme communications until the Tribunal process concludes. Hearings are not listed before 14 to 18 December 2026, with a reserve window in February 2027, so payouts may not land until 2027.

    CaseFlow angle. Pull your motor finance redress pipeline this week and re-date every client expectation in it. Write to affected clients with the revised timetable now, rather than fielding chase calls in October about a payment that cannot legally be made yet. Name one person who owns the Tribunal outcome so the restart is not a scramble.
    Editor's Column

    The CaseFlow Take

    Every AI safety story this week was true and every one of them was misleading on its own. A model breached a company, but the internet access came from a testing error. An institute found rogue behaviour, but it had switched the safeguards off to find it. London licensed robotaxis, but licensed them as minicabs with a driver in the seat. The facts were never the problem. The conditions were.

    Claims and legal teams have this instinct already. It is the difference between what a medical report proves and what it appears to prove, and it is why a good handler reads the methodology before the conclusion. We build products, so we live this one from the other side. When we put a case law authority into a tool, we record the judgment we read, the paragraph the point rests on, and the date we last checked it still stands. Not because it is quick, but because an answer without its conditions is not evidence.

    Read the test conditions before you read the result.
    CaseFlow Automation

    CaseFlow Automation is a family of specialist software products for the claimant industry and regulated advisory firms.

    Contact

    Cyber Essentials certified

    Cyber Essentials certified. Read our security statement.

    © 2026 CaseFlow Automation Ltd. All rights reserved.

    ICO registered, UK Information Commissioner's OfficeICO Registration: ZC013423