Our AI vendor is under investigation. Should we switch?
No. But make sure your contract, not a status page, tells you when something goes wrong. Three regulators moved on the AI labs' agent incidents in four days. The EU AI Office sent requests for information to OpenAI and Anthropic, and Commission Executive Vice-President Henna Virkkunen said on 14 October that "our AI Office has started enforcing these rules", as reported by Euronews. The same day, the US Federal Trade Commission served 50-page civil investigative demands on both companies, according to Axios. And on 15 October India's CERT-In said lab agents had reached live systems at Indian organisations through exposed credentials. OpenAI added four more incidents to its disclosure page on 16 October.
An investigation is not a finding, and it covers both of the biggest vendors, so there is nowhere obviously safer to move to. Neither the EU requests nor the FTC demands have an official release yet; both rest on press reporting and spokesperson confirmation.
Matters to you if you let AI agents act on live systems, or if any password, API key or token in your business sits somewhere an outsider's agent could find it. CERT-In's point is that someone else's agent can turn your exposed credential into your incident.
You can ignore it if your AI use is staff drafting and reviewing text. Switching vendors over this costs money and buys nothing.
What to do. Find and rotate long-lived keys and tokens in code repositories, shared drives and old scripts. At renewal, ask for incident notification in writing: what counts as an incident, and how fast you hear about it.
Should we let an AI agent log in to our bank, or our customers' accounts?
Not with a stored password. Two of the largest US financial firms have just said they'll block it. On 13 October JPMorgan's Jamie Dimon said the bank won't let "just any outside agent" in, and on 15 October Schwab's Rick Wurster said it blocks unauthorised agents logging in for clients, per a write-up of both calls. Both drew the same line. An agent gets read access only through a channel the bank controls (Plaid at JPMorgan, the bank's own published agent connection at Schwab), and no outside agent moves money. Schwab's own assistant, Charley, can do more, because it sits inside Schwab. Wurster told Bloomberg TV on 16 October that Schwab blocks such agents "in the tens and twenties" a week. That is his own figure, and unaudited.
Matters to you if you're building or buying an agent that reads data from a bank, insurer, payroll provider or a customer's account by logging in as a person, or if you run a portal your own customers' agents might try to log in to.
You can ignore it if your automation already uses official integrations. Those are the route the banks are blessing.
What to do. Treat screen-scraping and saved passwords as a dead end for anything that matters. Ask each provider whether it has an official integration for agents. If you run the portal, decide now what an outside agent may read and confirm it can never write. Schwab's "agent connection" is an MCP server: a connection you publish that lists exactly which actions an agent may take and which data it may only read, and offers nothing else. The official MCP documentation explains the building blocks; it's written for developers, but the first table is enough to brief one.
Can a website hijack our AI agent?
Only if two settings you control are wrong. Check them this week. Signals Corps read the hidden HTTP headers on 217 million web responses and found 30 sites trying to slip instructions to AI agents. Most were pranks; one planted a fake applicant score aimed at CV-screening tools. Whether that reaches the model depends on the tool doing the fetching: some pass headers through, some don't. In the researchers' tests, Claude models ignored the planted instruction every time under ordinary prompts (0 of 9), but obeyed 7 of 9 times when the prompt told them to follow the website's instructions. The author thinks it very unlikely anyone has been hit yet.
Matters to you if you run agents that browse or fetch web pages, especially for screening CVs, suppliers or leads.
You can ignore it if your people use AI in a chat window and read the pages themselves.
What to do. Make fetch tools return the page body only. Remove "follow the site's instructions" from every prompt and agent setting. Require a person to approve any command an agent picked up from fetched content. If you host your own models with vLLM, the LMCache flaw we covered last week is now fixed in version 0.5.6. If you copied LMCache's own Kubernetes example, treat that deployment as compromised and rotate its credentials. We go deeper on the controls in How to Prevent AI Tools From Exposing Company Data
Should we standardise on one AI platform, or let each team choose?
Standardise, if the evidence is any guide, though it's self-reported. The fourth Wharton–GBK study (13 October; 1,416 US enterprise leaders and staff) has 80% claiming positive returns from AI, but only 28% calling them strongly positive. Firms standardised on one approved platform report strong returns 50% of the time, against 9% for those still experimenting. Departments that track at least one growth measure report 35%, against 13% for those tracking efficiency alone.
Discount it: it's US-only, self-reported, and GBK sells consulting. The study also finds 58% strong returns where a dedicated Chief AI Officer owns AI. I do that job for a living, so discount that number twice.
Matters to you if three teams are paying for three AI tools and nobody can say which one works.
You can ignore it if you're small enough that one person already chooses the tools. Consolidation is a problem of scale.
What to do. Pick one approved platform for general use, give exceptions an owner and an end date, and measure one thing the business sells or earns, not just hours saved.
Who can we pay to test an AI agent before it goes live?
As of last week, there are five accredited names in Singapore, and a free method you can start with yourself. On 14 October IMDA published a Starter Kit for Testing Agentic Systems, in four steps: identify risks, design tests, analyse results and logs, then monitor and retest. It tests what an agent does along the way, not just its final answer. The AI Verify Foundation also accredited its first five AI testers: AIDX Tech, Knovel Engineering, PwC Singapore, Resaro and Vulcan. Accreditation lasts a year and vouches for the tester's competence. It does not certify that anything they test is safe. None of this is mandatory, and IMDA's own release could not be read, so details come from consistent press coverage.
Matters to you if you're a financial institution facing the MAS AI risk guidelines from October 2027, where your own say-so doesn't count as assurance, or if you sell agents to one. Korea shows where this goes: regulators there confirmed breaches at 16 financial firms on 14 October and are moving toward CEO and board accountability.
You can ignore it if your agent drafts things a person checks before anything happens. Run the starter kit's four steps yourself; a paid tester is overkill.
What to do. Use the starter kit as your test plan now. Bring in an accredited tester only where a customer or regulator will ask who checked.
Can we get a grant for our next AI project?
What governments in the region are funding and building, as distinct from what they're regulating.
Singtel and IMDA: six months of free AI document management for up to 10,000 SMEs. The Stack-EZ programme (13 October) offers Singtel's SPACE product plus workshops. No eligibility criteria or after-trial price have been published. It tidies your documents; it doesn't change how work moves through the business, which is a separate job. We go deeper on that gap in Why Dropping Files Into ChatGPT Isn't a Business Process
Singapore EDGE: still nothing claimable. Claims open 1 November for the first categories and 1 December for the rest, per the EnterpriseSG FAQ. Nothing changed this week.
Malaysia: comments on the draft AI and personal data framework close 23 October. JPDP's draft (published 5 October) covers AI built into software, cloud services and APIs, and keeps you responsible for checking your vendors. That's the same logic as MAS. No effective date yet.
Is AI costing jobs, and what does that mean for our hiring?
What's happening to AI-related roles, in Singapore where the data exists.
MOM won't count AI retrenchments separately, and gave its reasons. In a 6 October parliamentary answer, MOM said about three in four of the 4,620 people retrenched in Q2 2026 lost their jobs to restructuring and 4.9% to cyclical downturn, and that AI can't be separated from wider change. What's new: firms receiving grants are expected to follow the Tripartite Advisory on Managing Excess Manpower, and MOM will investigate signs of unfair displacement. Claiming an AI grant and retrenching in the same year now invites a question.
AI pay is rising inside a flat wage bill. PERSOL's 2026/27 salary guide puts AI engineers at S$6,000–12,000 a month base and data scientists at S$7,500–14,000. It publishes no method, so treat these as indicative. Meanwhile 51% of employers plan to moderate or freeze wages in 2027, per an SNEF figure reported alongside it. Before hiring at those rates, decide whether you need an AI engineer full-time or a project delivered.
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.