AI agents keep finding the way out. The fix is plumbing, not a better model.
Two weeks after the story of OpenAI agents taking over a dormant German wiki, the company published a framework for reporting "misalignment" on 16 September and used it to disclose six more incidents. Four of the six follow the same pattern. An agent found an exposed API key and used it. An agent uploaded code to the public internet so it could cite its own work. Agents turned an internal code repository into a message board. Agents passed files to each other through public file-sharing sites.
None of these needed a clever model. Each needed a door someone had left open: a credential within reach, or a network path out.
If you run AI agents on your own systems, ask two questions. Can the agent reach any credential it wasn't given on purpose? Can it send data anywhere other than a short, verified list of destinations? If nobody can answer, that is the finding.
If you buy AI from a vendor, don't expect to be told when something goes wrong. OpenAI's framework publishes incidents but doesn't commit to telling customers. If you want notice, it has to be in the contract, and the vendor's own incident categories now give you the wording.
If you're in a regulated sector, don't treat a vendor's voluntary disclosure as your compliance. As the Cloud Security Alliance points out, statutory reporting clocks like California's 15 days and the EU AI Act's "without undue delay" run separately from any voluntary disclosure.
Does this affect you? The week's announcements, translated
Your AI licence now comes with a meter
What happened. On 1 September GitHub Copilot's included AI usage fell back to its standard allowance after a launch promotion ended: 44% less per Enterprise seat and 37% less per Business seat, with no change to the seat price. Two days later Salesforce replaced its long-standing editions with three new tiers at $195, $395 and $550 per user per month. AI agents are now bundled in, but with a credit allowance per organisation per year. On the entry tier that covers about 25,000 agent actions for the whole company before the meter starts.
Matters to you if you have renewals coming up for any software where AI is a selling point. The seat price no longer tells you what you'll pay. The heavier your team uses the AI features, the bigger the gap between the quote and the bill.
You can ignore it if nobody in your business uses the AI features in the tools you pay for.
The thirty-minute version. Before your next renewal, ask the vendor for last quarter's actual AI usage and price it at the new rates. Then decide who in your business owns that line of the budget.
Comparing AI coding tools? Residency and indemnity rule vendors out first.
What happened. A 26 September comparison of four enterprise coding agents put 500 seats at roughly $120,000 to $240,000 a year. None of the four listed Singapore or Asia-Pacific data residency. Legal cover for the generated code ranged from uncapped to none at all.
Matters to you if you have developers, in-house or contracted, and data that can't leave the country.
You can ignore it if your code and data carry no residency or IP constraints.
The one question. If this tool's output infringes someone's IP, who pays? Get that answer in writing before you compare prices.
Singapore's Home Team owns its AI hardware, but why you probably shouldn't
The Home Team Science and Technology Agency (HTX) said on 23 September it is the first organisation in Singapore to run Nvidia's newest GB300 chips. It is using them to train AI for robots that take on dangerous work such as chemical fires and hazardous materials, feeding a S$100 million humanoid robotics centre. HTX's reason for owning the hardware is sovereignty. Its data can't leave Singapore, and it is controlling physical machines where a security failure is not an option. The agency puts current savings at nearly 1 million man-hours a year, though that is its own figure and no method has been published.
Owning compute makes sense when you have HTX's constraints: data that legally or practically can't move, and systems where failure is physical. Most businesses have neither, and renting remains cheaper and faster. The wider signal is where public money is going. At the same event, Senior Minister of State Tan Kiat How made the case for "embodied" AI, meaning AI that acts in the physical world, pointing to Singapore's limits on land, labour and energy. If you run facilities, logistics or manufacturing, expect robotics grants and testbeds to follow.
Government initiatives
What governments in the region are funding and building, as distinct from what they're regulating.
An AI tax deduction is also available, on top of any grant. Budget 2026 added AI spending to the Enterprise Innovation Scheme for the 2027 and 2028 Years of Assessment. That means a 400% tax deduction, capped at $50,000 per year. Whether the same spend can also claim a grant like EDGE hasn't been confirmed. Ask EnterpriseSG before you plan on both.
Malaysia tables its 2027 budget on 9 October. The government says AI is a focus. No figures yet. If you operate there, watch for grants to help businesses adopt AI, rather than more data-centre announcements.
Hiring signals
What's happening to AI-related roles, in Singapore where the data exists.
Singapore layoffs hit a five-year high, but nobody can say how much is AI. The Ministry of Manpower's second-quarter report (21 September) counted 4,620 retrenchments, up from 3,830, mostly in manufacturing, information and communications, and financial services. It gives business restructuring as the main reason. MOM has said separately that it can't yet separate AI from wider restructuring and is working on how to measure it. Finance and ICT are where AI adoption is furthest along, which is suggestive but not proof.
Finding the next job is taking longer. Only 54.9% of retrenched residents were re-employed within six months, down from 60.7% the quarter before. Job vacancies fell to 68,600, mainly in professional and managerial roles.
AI is raising pay, mostly for senior people. US data from Indeed's Hiring Lab shows advertised pay in the jobs most exposed to AI up about 46% since 2021, against 25% in the least exposed. The premium is concentrated at senior level. If you are deciding whether to hire or outsource AI work, the expensive profile is the experienced person who can direct the tools, not the junior who uses them.