Can the new PHP security flaw wait for our next scheduled update?
No, not for anything on the internet that runs PHP. Two critical flaws in the PHP engine itself, each rated 9.2 out of 10, let an outsider take over a server without logging in. Clean application code doesn't help, because the flaw sits underneath it. Exploitation attempts were reported about twenty minutes after the patches shipped. BambooHR and Dutch government services including DigiD went offline as a precaution; neither reported a breach. The PHP Foundation also admits a developer saw signs of one flaw in a 2024 audit and the fix stalled, calling that "a structural failure" of its own process.
Matters to you if you run any PHP: WordPress, a custom ERP or CRM, a forgotten marketing site.
You can ignore it if you are certain nothing you run uses PHP. Confirm that with your host rather than assuming it.
What to do. Ask your host or developer which PHP version you run and whether it carries the fix: 8.5.12, 8.4.27, 8.3.36 or 8.2.35 or later. Some Linux systems patch without changing the version number, so ask them to check the package changelog for the two CVE numbers (CVE-2026-109618 and -109619). Setups using Apache's mod_php are the riskiest. PHP 8.2 stops receiving fixes on 31 December 2026. We go deeper on this in our PHP2Shell write-up.
Are we too small for anyone to bother attacking?
No. One person's AI agents broke into sites at more than 100 companies for about $50 each. Researchers at NorForge found the attacker's exposed control panels, and Reuters used his own records to name victims, including Deere, SharkNinja and Cummins, according to coverage of the findings. About one attempt in ten worked, each taking nearly 23 hours, on roughly $5,000 of AI usage. What got hit were support and marketing sites, not core systems. The figures are the researchers', and we haven't seen the underlying records.
Matters to you if you have any website or login portal that nobody specifically owns.
You can ignore it if everything you run sits on a platform whose provider handles patching. Even then, check your own passwords.
What to do. List every site, portal and login that faces the internet, with one named owner each, and retire what has none. Rotate passwords and keys left in old code folders or shared drives.
Could a routine software update break our AI tools?
Yes, and this week it can happen silently. Anthropic's new API version removes the temperature, top_p and top_k settings. Sending any of them now returns an error, and the newest software kits send the new version by default, so a routine upgrade is enough to start the errors. If your developers lowered temperature to get consistent results, that effect is gone and Anthropic says no other setting replaces it. Separately, OpenAI retired older models on 23 October, along with every fine-tune built on them. And a compromised release of LiteLLM, a popular gateway that sits in front of AI providers, collected cloud keys and AI-assistant credentials from machines that installed it.
Matters to you if a developer or vendor built something that calls these services, especially anything extracting or classifying data. Also if you run a LiteLLM gateway.
You can ignore it if your staff only use the chat apps directly.
What to do. Ask your developer to pin the software version and search for those three settings before upgrading. If you run LiteLLM, go back to version 1.106.2 or rebuild, and rotate every key it could see. A fine-tune only lives as long as the model it was built on, so don't treat one as an asset you own.
Our vendor says AI saves them 20%. Why aren't we seeing that?
Because that figure is usually the vendor measuring its own product on itself. SAP says its assistant gives about 20% productivity gains across 110,000 of its own staff, on its earnings call. In the same results its cloud margin fell 0.7 points and services revenue fell 5%. IBM says AI is about 30% of its consulting revenue, while consulting grew 2%. Barclays is cutting about 5,000 roles, yet its CEO called AI "the icing on the cake, not the cake".
Matters to you if someone is selling you AI on a promised percentage saving.
You can ignore it if you are buying a specific feature you will use, not a saving.
What to do. Ask who measured the saving, against what baseline, and at what cost. Then test one workflow with your own numbers before paying for more.
Does the EU AI Act apply to us yet?
Only if you sell into the EU. Italy is the first country to have enforcement powers in force. Its AI Act decree took effect on 23 October, per a legal analysis: the national cybersecurity agency can inspect AI providers without notice, with fines up to €35 million or 7% of worldwide turnover. The next EU date is 2 December 2026, when AI-generated content must be marked.
You can ignore it if you have no customers in the EU. What to do if you do: confirm whether your product generates synthetic text, images or audio.
Can we get a grant for our next AI project?
What governments in the region are funding, as distinct from what they're regulating.
Singapore: more than S$45 million over three years for AI in manufacturing. EnterpriseSG will offer small manufacturers a suite of ready-to-deploy AI solutions: component design, quotation-document processing, production scheduling and predictive maintenance. There is no published co-funding rate, cap, eligibility or sign-up route yet, and nothing says who supplies the solutions, so ask EnterpriseSG before budgeting around it. Not for you if you aren't a manufacturer; press reports say a logistics and wholesale version is in development.
Singapore EDGE: still nothing claimable. The EnterpriseSG FAQ is still pre-launch. Claim dates of 1 November and 1 December appear only in advisers' write-ups, which disagree on the second, so don't plan around them.
Malaysia: comments on the draft AI and data protection framework closed 23 October. It has no compliance deadlines, and industry's main ask was a 12 to 18 month grace period. Plan for 2027.
Is AI changing who we should hire?
What's happening to AI-related roles, in Singapore where the data exists.
Employers are asking for AI engineers, not data scientists. Jobstreet's new AI Gauge puts Singapore job ads mentioning AI skills at 5.1% in August 2026, up from 1.3% in January 2024. "AI engineer" and "automation engineer" have displaced "data scientist" in the top three titles. No method is published, and about 95% of ads still mention no AI.
Nobody owns the work. A vendor-commissioned US and Canadian survey found 45% of mid-sized firms have no single owner for how people and AI work together. The sponsor sells exactly that role, so treat it as indicative. Before hiring an engineer, name an owner for the outcome. If you can't, a hire won't fix it.
Big employers are cutting about 5% without crediting AI. Barclays and American Express both announced this month, and both framed it as cost, not AI. That is anecdote, not a trend line.
What's your question about AI in your operations? Reply to this email or send it through bitnatives.com/contact, and we'll answer it in a future issue.