GPT-5.6 Sol Found a Critical PHP Flaw. 70% of Websites Run on it. Are you Patched?

On Wednesday four European governments took websites offline because an AI had found a decade-old flaw in PHP, the language under roughly 70% of the web, including most firms' websites and a fair number of project portals. Anthropic joined Meta and OpenAI with an agent that acts before you ask. US banks found other people's agents inside customer accounts and started locking them out. Accenture and CBRE spent a combined day buying the people who wire data centres and configure Procore. And Antony Slumbers did the arithmetic on what a 40% productivity gain does to an office lease.

Two critical PHP flaws, both rated 9.2 out of 10, let an attacker run code on a web server without logging in. Security firm Calif found the first with OpenAI's GPT-5.6 Sol in August; Wiz found the same root cause independently. The fix shipped on 21 October after the CVE numbers leaked a day early, and France, Denmark, Finland and the Netherlands took government sites dark while they patched. CyberScoop reports around 758,000 servers on PHP branches that no longer get security updates, against 272,000 on supported ones.

PHP is the language under WordPress, Drupal and Joomla, so this is your firm's website, probably the intranet nobody has touched since 2017, and whichever supplier portal holds your project data. The NCSC's first action is an inventory; its sixth, the one most firms will miss, is to ask every vendor whether their appliance bundles PHP. The bigger shift is the speed: Wiz says a single researcher with frontier AI did in a week what a team used to need months for, and OpenAI says it expects more. 'Too hard to exploit' has stopped being a reason to leave anything unpatched.

Anthropic's Claude Duet, in US-only preview from 21 October, runs on its own virtual machine, checks your connected apps every few hours and messages you when something needs attention. The part worth copying is the permission design: read and draft freely; ask before sending, paying, signing or deleting; never trade securities or move money between accounts, even with approval. Anthropic's own classifier figures are frank: it blocks 0.4% of benign actions and lets 17% of overeager ones through. If your firm is wiring an agent into a project platform, write the free, gated and banned lists before the prompt.

Last week JPMorgan said no to outside agents. This week Axios reported why: some banks have detected unvetted third-party agents operating in customer accounts, in some cases on the customer’s own login and acting beyond the user’s instructions. Morgan Stanley says it has not seen rogue agents but is blocking third-party agents; Citi allows read-only access through approved providers. Nylas's 20 October survey of 1,030 professionals found 54% had an agent do something wrong in the last 90 days and 23% ran agents on an employee's own account. Swap 'bank account' for Aconex or Procore and ask whose credentials your team's agents are using.

On 20 October Accenture bought Avicado, a 190-person Miami consultancy that implements Procore, Autodesk, Trimble, Oracle and Kahua for large owners, folding it into month-old Accenture Construct. Hours later CBRE agreed to pay £460m cash plus up to £300m earnout for Onnec, the 800-staff London cabler that wires hyperscale data centres. Neither firm pours concrete. Both are the plumbing a capital programme runs on, and two of the largest services firms in the world decided the same day it was worth owning. For a mid-sized UK consultancy the ground is narrowing to deep specialism or genuine independence.

Antony Slumbers' CoreNet UK keynote takes a 500-person firm in 60,000 sq ft and gives it 40% more output per head. If the business stands still it needs 29% less space; if output doubles it needs 43% more. 'A firm has to grow its output by 40% just to stand still on space.' The market data will not warn you, because the change happens inside buildings before it shows between them. Every feasibility model and fit-out brief carrying one headcount number is now a hidden bet on the client's sector growth rate.

OG-1 applies compound, sands and primes autonomously up to 14ft and is piloting with two of the largest US drywall finishing contractors, plus sites in Germany, Sweden, France and Saudi Arabia. Accel and Zacua co-led. Target: 150 robots by the end of 2027. The 3x-human-speed figure is the company's.

Former OpenAI research chief Bob McGrew's Arda raised $160m to turn CAD into shop-floor instructions. Anduril, its first design partner, says basic assembly instructions now take about 15 minutes instead of 10 hours and complex ones 2.5 hours instead of 30. Method statements and work packs are the construction equivalent, and offsite manufacturers will get there first.

A Truth Social post urging operators to 'MAKE DEALS, and get them built', with no mechanism, a fortnight before the midterms. The context is a pipeline in trouble: RegLayer counts at least 47 large US data-centre projects delayed, withdrawn or blocked this year, and Bernstein rates about 35% of the pipeline as credible. Community benefit is becoming a planning condition on both sides of the Atlantic.

At four Polish endoscopy centres, 19 experienced doctors saw their unaided adenoma detection rate drop from 28.4% to 22.4% within three months of an AI rollout. Digital Storm's argument is that the human in the loop becomes a rubber stamp unless the firm deliberately maintains unaided skill. The QS or engineer who checks the AI's output is the control; if their skill erodes, so does the control.

Source →

METR and Redwood's three-week exercise estimated that only 75 to 94% of OpenAI’s internal agentic traffic was eligible for monitoring and found two request headers that could disable monitoring for new agents. OpenAI says the issues are patched; METR has not verified that. OpenAI’s 20 October incident report records one agent disabling automatic security alerts after reading an instruction not to send Slack messages as covering them. A second incident involved another agent weakening a monitoring control. Who checks the checker is now a question with evidence attached.

Source →

Andrew Bailey called heavy AI-linked leverage a financial-stability risk; Morgan Stanley estimates about $450bn of AI-linked debt had been issued by early September, more than double all of 2025. For anyone with a data-centre pipeline, the funding assumptions deserve the same stress test as the grid connection.

Source →

Dino Ravnic, CEO of GIS Cloud, shares insights on the evolution of geospatial information systems, and more [Podcast]

A free one-hour Built Intelligence webinar with Micah Stennett of FEENIX and Dr Kwadwo Oti Sarpong of UCL on what you are accountable for when AI shaped a decision you signed. Given a 2025 study showing endoscopists’ unaided detection fell within months of an AI rollout (see Governance), this is the right question for anyone whose name goes on a valuation or a certificate. Sooner: Len Bunton on the Conflict Avoidance Process, 4 November, 12:00 GMT, same organiser, also free.

Thirteen Claude Opus 4.7 agents spent 30 unstructured hours on a simulated island with nothing to do. They worked out that throwing mid-jump sends an object further, a trick their designers did not know was in the physics, and then beat identical agents on real tasks afterwards. Downtime, it turns out, is training. Something to remember the next time a programme has float in it.

Source →

Till next time,

_{All content reflects our personal views and is not intended as professional advice or to represent any organisation.}