Forge Signals | October 30, 2026 | Editorial
The screen we no longer need
Google’s Android documentation begins with a reminder. A user asks an assistant to remember an errand, and the assistant invokes a task application’s function to create the entry. The person never opens the app. They never see its form or press its Save button. The convenience is real. So is the change in where the work happens. The application is now serving a caller that may not be the person whose information it holds. [1][2]
At Android Dev Summit on October 28, Google announced that AppFunctions had reached beta. Developers can expose selected app capabilities to compatible agents through Android MCP, while the separate Connected Apps Platform preview offers integration with Google products including Gemini and Search. Google’s presentation also introduced an interface renderer and on-device agent APIs at an earlier alpha stage. These are not identical products, and the preview labels matter. Google is designing several routes through which machine callers can use software built around human interaction. [1][3]
A declared function with a schema gives an operator a much clearer contract than a model improvising through a changing screen. Android requires callers to hold permission to discover and execute AppFunctions. Connected Apps documentation gives developers ways to specify their tools, scopes, authentication and confirmation requirements. Those permission mechanisms are real. They also operate at the level of the caller and the function, which may leave the app blind to the larger task that triggered the call. A reminder might safely tolerate that gap. A change to a payment destination, an account setting or a file shared with another service might not. [2][3]
A screen is an awkward interface for automation, but it often carries the point at which a user can see what is about to happen and refuse. Developers who remove the screen from the operating path inherit the work it was doing. They must decide which operations need a separate confirmation, what context the caller must supply, and how to preserve evidence of the user’s intended action. Google’s design now puts those decisions directly in the developer’s path. Their effectiveness will have to be demonstrated in the applications that use it.
The warning GitHub cannot enforce
A malicious instruction embedded in a web page can become dangerous when an agent reads that page while holding access to tools. The model may treat words it was supposed to analyze as directions it should obey. GitHub’s October 28 announcement offers code scanning intended to find vulnerable paths before an AI application ships. Its public-preview release describes indirect prompt injection, exposure of sensitive information and excessive agency as problems worth addressing. [4]
The detailed documentation draws a tighter perimeter. For now, the scanner detects indirect prompt injection only. The scanner covers new pull requests. Existing repository code receives no equivalent review from this feature, and pull requests from forks and Dependabot are excluded. An eligible organization must enable it. GitHub Advanced Security is required, with AI credits consumed by the scans. That narrower coverage governs what the scanner can find today. [5]
GitHub says the findings are advisory comments, and its documentation does not permit enforcing them as a merge requirement through a ruleset. An engineer can read the warning, investigate it and decide the risk is acceptable. Another can dismiss it to meet a release deadline. The preview provides evidence for the decision, while the decision and its consequences remain with the people controlling the release. An advisory scanner may be the right place for a new product to begin. A management report claiming “agent risks controlled” on the strength of its installation would still be indefensible. The team owns the merge decision, and GitHub’s documentation makes that limitation explicit. [5] A pull request captures one version of the code. After deployment, the agent may encounter new retrieved material and operate with permissions the reviewer never tested. The scanning result belongs in a release decision that accounts for those runtime conditions.
The bank will not accept the agent’s password
Royal Bank of Canada drew a line on October 27. It announced that third-party personal AI agents would be blocked from direct access to authenticated RBC digital properties, including online banking. The Canadian Press subsequently reported restrictions or non-support positions at TD, BMO, CIBC and Scotiabank. Their policies differ, so there is no basis for describing the news as one coordinated prohibition. [6][7]
Banks can authenticate a session without knowing whether the customer authorized the particular action an agent is about to take. A customer might ask for help paying bills, but that request does not by itself authorize every transfer a model could infer from the task. Password sharing makes the ambiguity worse because the bank may see an authenticated customer session where the customer sees a delegated machine action. RBC cited fraud, consumer harm and wider financial risks in explaining its decision. Its notice does not establish that any of the agents it excluded had already caused the systemic harm it fears. [6]
Refusing access is an enforceable decision within the bank’s own systems. It also leaves the customer without the delegated service they wanted. Someone who wants an assistant to assemble account information across institutions has a legitimate interest in doing so safely. The bank has a legitimate interest in refusing a third party it cannot identify or constrain. The interests conflict at the point where a customer wants delegation the bank cannot reliably distinguish from account takeover. What would resolve the tension is a delegation mechanism with separately identifiable agents, limited powers, transaction-specific evidence and revocation that works without sharing a password. The public announcements do not establish that such a mechanism is available for the excluded agents.
A further question deserves scrutiny without pretending to know the answer. Banks control valuable interfaces to customer financial activity, and restricting third-party intermediaries also preserves that control. Security may be the entire reason for the decisions. Commercial interests may be involved as well. The published statements establish the stated security rationale, not the institutions’ private motives. In the meantime, customers and agents know what the bank will refuse. They have much less clarity about the conditions under which the bank would permit a safer form of delegation.
The successful response that leaked a stranger’s data
On October 26, Cloudflare published an account of a confidentiality failure in its GLM-5.3 model on Workers AI. From August 27 to September 9, the company identified 155 responses containing information from another customer’s request. All involved prompts longer than 681,280 tokens. Cloudflare reported no evidence of deliberate exploitation, and its investigation found that no other Workers AI model was affected. These are findings from the operator’s investigation, supported by an unusually detailed account of the failure path. [8]
The model offered a context window of 1,048,576 tokens, but the key-value cache on its prompt-processing GPUs could hold only 681,280. Cloudflare knew the capacities did not match when it launched the model. It expected requests exceeding the available cache to fail with an error. Its long-context tests returned successful responses instead. The company had checked status and performance without verifying that the returned content was coherent or belonged to the request. The tests passed because they were testing the wrong thing. [8]
The rejected request reached a dangerous handoff in SGLang, the inference engine. SGLang marked the oversized input for rejection and replaced it with a one-token placeholder, but the distributed request continued into generation. The receiving GPU was told that the cache transfer had finished even though memory pages from previous requests remained. Generation could then attend to those pages and include another tenant’s information in the answer. The client received HTTP 200. Its own request appeared to have succeeded while the service failed tenant isolation. In a shared inference service, one customer’s data must never appear in another customer’s answer. [8]
There was no rogue agent to restrain. The failure sat in the operator’s memory configuration and in the engine’s handling of a rejected request. That matters because the error survived a control the service had already attempted to exercise. The request was marked for rejection. Somewhere between that decision and the generation stage, the machinery treated incomplete work as completed work. A monitoring system that celebrated HTTP 200 would have reported success throughout.
Cloudflare reproduced the problem, disabled the affected model, expanded the available cache, repaired the rejection path and tested the changes before restoring service. Its disclosure explains the work with SGLang maintainers and the additional hardening applied to related deployments. The remediation is concrete, and so is the limit of what it can establish. Cloudflare can prevent this known path from exposing another customer’s data again. It cannot retrieve text that has already appeared in someone else’s response. The published account does not show what recipients did with that text. We should neither invent downstream misuse nor confuse the absence of evidence for misuse with recovery of the information. [8]
The electricity connection belongs to somebody
Two days after Cloudflare’s disclosure, the British government published its Data Centre Covenant. The document begins from a strategic ambition to host more computing infrastructure in Britain. The government estimates that meeting its objectives will require 6 gigawatts of domestic AI data-centre capacity by 2030. That is the government’s estimate for planning and policy. It is not an engineering fact the document independently proves. [9]
The covenant asks operators to carry a fair share of energy-system costs, invest in host communities, account for water use and strengthen security. Developments seeking additional government support as strategic projects would face stronger expectations, including local benefit plans and annual reporting on energy and water. In exchange, the government proposes measures to help qualifying projects through planning and electricity-connection constraints. The government has not yet published all the criteria for strategic designation. The bargain has been proposed. The results have not yet been measured. [9]
A developer requests electricity at a specific site. The network operator has to consider the homes and businesses already connected, together with other projects waiting for capacity. A faster connection for one facility can change the timetable for other applicants when the network is constrained. Local jobs and investment may justify that priority, or fall short of what the host community ultimately pays. The government argues that new demand could spread fixed electricity costs across more users. Whether households save money depends on investment, tariffs and measured demand. The covenant cannot settle those questions in advance. [9]
This policy is an attempt to make a private demand for computing capacity answerable to public obligations. Specific projects will reveal whether the covenant’s terms have practical force. A community needs to know what was offered in return for hosting a facility, how much water and electricity it consumes, what benefits were delivered, and who can challenge the accounting. The publication establishes expectations for future decisions. It does not yet establish whether the people carrying the costs will receive the promised return.
After the control has done its job
Only one of these five developments documents an actual breach. Google’s interfaces are being introduced, GitHub’s detector is in preview, and Canada’s banks have made access decisions. Britain’s covenant sets terms for projects that have yet to demonstrate outcomes under those terms. Cloudflare alone supplies a confirmed confidentiality breach. That difference is central to what the evidence can support. An operating model under construction and an incident that already happened require different claims.
An organization can complete every action within its own reach and still leave obligations outside it. GitHub can document a warning without preventing a merge. Cloudflare can repair an inference engine after a customer’s data has already crossed into a stranger’s answer. The records can be accurate while the exposure persists or the next decision takes place beyond the original operator’s visibility. That is where the ordinary vocabulary of control begins to lose precision.
There is an emerging separation between the ability to control an AI operation and the ability to contain its consequences. We should be careful about the first half of that sentence. Operational AI governance is not a solved problem that has merely developed an awkward edge case. Some controls are experimental, some are voluntary, some are advisory, and a known capacity mismatch was allowed into production on the assumption that an error path would behave correctly. Having a framework, a scanner or a permission model establishes the existence of a mechanism. Its effectiveness has to be demonstrated against the property it was built to protect.
Consequences outlasting the original operation are hardly new. An unauthorized transfer can have consequences after credentials are revoked. An exposed document does not return home when a vulnerability is patched. Utility decisions create obligations lasting longer than the project that prompted them. The difference AI may introduce is the ease with which one interpreted instruction can initiate further work through systems administered by different owners. That effect is deployment-specific, not some universal property of every model. The risk presented by an assistant drafting a paragraph is nothing like the risk presented by an agent changing production infrastructure or moving money.
Our governance reporting is far more comfortable recording permission than tracking aftermath. It can say who approved an action and whether a check ran. It rarely shows the state changed in a different service, what a recipient did with exposed information or whether the promised savings from a grid allocation reached households. The records may each be correct. An obligation can still go missing between them.
The accounting nobody wants to own
Cloudflare offers the cleanest place to see the problem because its disclosure is unusually specific. An operator can show the corrected cache configuration, the repaired rejection path and the tests used to verify them. Those are meaningful pieces of evidence. The operator cannot show, from that technical repair, that every recipient of an affected response did nothing further with the information. The disclosure makes no such claim. Whether notification or remedy is owed depends on the exposure and the applicable duties. Restoring the model cannot settle those questions by itself.
A flattering governance report is easy to assemble from the other announcements. A team enables GitHub’s scanner and records that code scanning is active. An application exposes functions behind a valid permission check. A developer signs an agreement promising local benefits for a future data centre. Each entry can be accurate. The dishonesty begins if those entries are offered as evidence that the risks they address have been contained. GitHub’s own instructions say its findings cannot prevent a merge. The Android caller permission cannot prove the user’s intent for an entire workflow. A signed promise cannot establish whether a neighborhood received what it was owed years later.
The accounting breaks when evidence ends at the operator’s boundary. It can explain what its system allowed, blocked, inspected or repaired. Someone else may be left establishing the cost of that decision after the operator has moved on. The people who receive the consequence are often least able to reconstruct the machinery that produced it. A customer cannot inspect the cloud provider’s cache transfer. A resident cannot infer a fair grid allocation from the ceremonial signing of a covenant.
No responsible engineer can promise to predict or reverse every downstream effect. An honest governance claim begins with the reachable effects the operator can identify, the evidence it will retain, the failure modes it has tested and the obligations that persist when stopping the system is no longer enough. Unknown downstream effects belong in the original authorization decision. They should limit the authority granted where the stakes warrant it, and they should remain visible after deployment. Calling them out of scope may simplify an audit. It does nothing for the person still living with the result.
We can build useful AI systems without pretending that completing an operation completes our responsibility for it. But the claim of governance has to survive contact with the people and systems downstream. If the only proof of success is that the tool ran, the request returned or the ticket closed, the operator has accounted for its machinery. It has yet to account for what that machinery did to anyone else.
Artifacts are cheap, judgement is scarce.
Per ignem, veritas.
Sources
[1] Matthew McCullough, “Agentic and Flexible: A next era of Android development,” Android Developers Blog, Oct. 28, 2026.
[2] Google, “Overview of AppFunctions,” Android Developers, accessed Oct. 30, 2026.
[3] Google, “Connected Apps Platform MCP server developer guide,” Google for Developers, updated Oct. 27, 2026.
[4] GitHub, “Code scanning for AI applications in public preview,” GitHub Changelog, Oct. 28, 2026.
[5] GitHub, “AI applications scanning for pull requests,” GitHub Enterprise Cloud Docs, accessed Oct. 30, 2026.
[6] Royal Bank of Canada, “Understanding third-party AI agents and your RBC banking,” Oct. 27, 2026.
[7] Daniel Johnson, “Some Canadian banks blocking personal AI agents amid security concerns,” The Canadian Press, Oct. 28, 2026.
[8] Kevin Flansburg and Michelle Chen, “Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang,” Cloudflare Blog, Oct. 26, 2026.
[9] HM Treasury, “The Data Centre Covenant,” GOV.UK, Oct. 28, 2026.








