In one minute
- Banks are restricting access by outside agents. Reuters confirms that banks are blocking them; alleged actions beyond the customer's instructions remain accounts from industry sources.
- 1Password and OpenAI expand the tools available to agents. One protects credentials; the other speeds up narrowly defined decisions. Neither replaces limits on the agent's authority.
- AI advances amid testing and physical constraints. Microsoft improves experimental diagnostic results. Nexans points to cable manufacturing capacity as a brake on data centres.
Lead story
Banks want to know who is entering and what they can do
FINANCE · AGENTS — Reporting from 22 October 2026.
An AI agent can carry out tasks as well as answer questions. Reuters confirmed with two sources that major US banks are blocking outside agents and transactions made through customer accounts. Its report followed an Axios investigation. [1][2]
Morgan Stanley has restricted such access, although it had not detected unauthorised or rogue agents, according to the source consulted by Reuters. JPMorgan had already explained on 13 October that it blocks services using shared credentials wherever it can. [2][3]
The key distinction: a customer may have permitted access even if the bank prohibits it. Accounts of actions apparently outside the customer's instructions come from industry sources; the reports reviewed provide no technical logs with which to assess their extent. [2]
Why it matters. Sharing a password leaves the scope of the task unresolved. Checking a balance, preparing a transfer and executing it are different permissions. Useful delegation must allow those permissions to be set and withdrawn, and the agent's actions to be reconstructed afterwards. That control matters as much as identifying the account holder.
Sources: 1 · Axios · 2 · Reuters, republished by MarketScreener · 3 · JPMorgan, 13 October transcript.
Other news
1. 1Password lets users authorise access without giving the agent their password
IDENTITY · SECURITY — Announcement on 22 October 2026.
1Password for Agents identifies the agent and asks the user to approve access from their device. It then enters the credential directly into the form without passing it to the model, according to the provider. Its partners include Meta, Anthropic and OpenAI. [4]
It is offered at no additional cost for Individual and Family accounts on 1Password.com; accounts on 1Password.eu and 1Password.ca are not supported. Integrations with Claude Duet and ChatGPT Work are still described as forthcoming. Each connection is optional, and availability depends on the partner. [5]
Keeping the password separate from the agent reduces credential exposure. But approving a login leaves another decision unresolved: what the agent may do once inside.
Sources: 4 · 1Password announcement · 5 · Availability and integrations.
2. OpenAI opens Decisions API for fast, narrowly defined decisions
SOFTWARE · AGENTS — Generally available since 22 October 2026.
Decisions API uses GPT-6 Luna to classify information, choose between options or assign scores. According to OpenAI, it can respond up to ten times faster than the same model through Responses API. The developer defines the question and the permitted answers. [6]
The base price is $0.10 per million input tokens, with no charge for output tokens. It supports up to 400 questions per request, structured data and integration with GPT-Live. Regional processing is available in the United States and Europe — the EEA and Switzerland — subject to specific conditions and surcharges. [6][7]
Box, Datadog, Ramp and Sierra are among the users named. The performance improvements are reported by the provider and its customers: they need to be assessed for each task. [6]
Sources: 6 · OpenAI announcement · 7 · Documentation and pricing.
3. A Microsoft system outperforms doctors in diagnostic tests
RESEARCH · HEALTH — Published in Science on 22 October 2026.
MAI-DxO organises a virtual team that proposes diagnoses, reviews hypotheses and decides what information to request. The research evaluates 304 complex cases from the New England Journal of Medicine and 325 emergency department episodes followed by hospital admission. [8][9][10]
On the complex cases, it achieved 83.2% accuracy, compared with 36.1% for generalist physicians using reference tools. On the emergency cases, it achieved 80.3%, compared with 46.8%. These are controlled case-based tests, not results from autonomous patient care. Microsoft says prospective studies are the next step. [8][9]
The news is the scientific publication and the expanded evaluation. The preliminary work was released in June 2025. The lesson is specific: organising the review of hypotheses can improve diagnosis. [8][11]
Sources: 8 · Microsoft · 9 · Science, authors' abstract on Crossref · 10 · Evaluation data · 11 · 2025 preprint.
4. Nexans warns of insufficient cable manufacturing capacity
INFRASTRUCTURE · ECONOMY — Statements from 22 October 2026.
Vincent Piquet, Nexans' chief financial officer, told Reuters that expanding the production of certain cables takes years. Building a line for cables with XLPE insulation takes around three years. That timeframe applies to this specific industrial process. [12]
Piquet said that Enedis, the French electricity distribution operator, has moved to six- or seven-year contracts to secure supply. Nexans has already announced a seven-year contract, including three optional years. This is a company's warning about industrial capacity: AI expansion also depends on components whose production cannot accelerate at the pace of software. [12][13]
Sources: 12 · Reuters, republished by The Star · 13 · Nexans, Enedis contract.
The signal in the noise
Autonomy needs permissions that can be verified
Today's editorial reading is straightforward: an agent's ability to complete a task must come with understandable authorisation. Accessing, deciding and executing are different steps. It helps to know who permits each one and how to stop it.
The advantage of a complete system will lie in connecting these pieces. A protected password, a fast response or a strong experimental result adds value when the limits of its authority are clear.
Spain and Norway
Spain: PHP flaws trigger a race to protect services
On 22 October, El País reported millions of exploitation attempts detected by Cloudflare and precautionary service shutdowns. These are attempts, not confirmed intrusions. The newspaper cited ministry sources who, at that time, were unaware of incidents affecting Spanish ministries and state bodies. [14]
INCIBE urges users to update PHP. If a system exposed to the internet cannot be updated, it recommends isolating it until a patched version is installed. The news is the response to vulnerabilities already disclosed. [15]
Sources: 14 · El País · 15 · INCIBE-CERT.
Norway: the sovereign wealth fund remains cautious about tech companies
The fund recorded a return of −0.5% in the third quarter and an accounting result of −278 billion Norwegian kroner. NBIM attributes the decline to the impact of rising interest rates on fixed income. [16]
Separately, Trond Grande told Reuters that the fund maintains small underweight positions in some large tech companies. He warned of AI volatility and a possible stock market setback if an incident slows adoption, without describing it as a systemic risk. This is investment caution, not a general withdrawal from AI. [17]
Sources: 16 · NBIM, results from 22 October · 17 · Reuters, republished by Euronext.
The AI Act remains under examination for incorporation into the EEA
EFTA's official record still lists Regulation (EU) 2024/1689 as under examination for incorporation into the EEA Agreement. Its application dates in the European Union do not automatically carry over into Norwegian law. [18]
Source: 18 · EFTA, AI Act status.