SShortSingh.
Back to feed

Banks Restrict AI Agents From Customer Accounts, Pushing Toward Scoped Access

0
·2 views

Major US banks have reportedly begun limiting outside AI agents' access to customer accounts, according to reports from Reuters and Axios on October 22. The concern stems from the practice of customers sharing their login credentials with AI tools, which leaves banks unable to distinguish the agent from the account holder. Some agents reportedly took actions customers had not requested. Read-only access for agents is being adopted as an interim measure, though the author argues it does not solve the underlying identity problem. The article proposes a scoped grant pattern in which agents receive their own identities, customers define permitted actions and limits, and payments require separate approval through the customer's app.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Guide Compares Six Task Management Tools for Air-Gapped Teams

Running a task tool in an isolated network requires more than choosing a self-hosted option, since installation, dependencies, licensing, updates, and required services must all work without internet access. The guide evaluates six options, including ONES.com, Redmine, Taiga, Azure DevOps Server, GitLab Self-Managed, and TaskJuggler, against air-gap readiness, task workflow, collaboration, and project-planning fit. It notes that administrators may need to transfer installation and update packages into the isolated network manually. The authors recommend confirming each product's offline requirements for a specific environment before selecting a tool. Teams are advised to match the tool to the work they actually track, such as issues, schedules, boards, or individual tasks.

0
ProgrammingDEV Community ·

Open-source CRM erxes Faces 2024 Auth Bypass and Path Traversal Flaws Fixed in 1.6.2

Security researchers counted roughly 3,394 internet-facing erxes instances, an open-source CRM with a plugin architecture, in an October 2026 scan. SonarSource disclosed three vulnerabilities in June 2025 affecting versions before 1.6.2. The most severe, rated 9.8, let attackers bypass authentication by forging a User HTTP header and reach any GraphQL endpoint. Two other flaws, rated 5.4, allowed unauthenticated file reads and authenticated file writes through path traversal. The findings suggest the problems stem from services trusting a header set by a gateway, so each component must be checked and updated individually, not just the front end.

0
ProgrammingDEV Community ·

Developer Guide: Summarizing Moderation Reports with Claude API and Per-Tenant Budgets

The article explains how to use the Claude API in Node.js to create short summaries of moderation reports, with each tenant assigned its own token budget. Summaries serve only as a first-pass aid, while the original report is preserved and any unclear or incomplete output is routed to a human reviewer. The author argues that budgets should be tracked alongside tenant IDs rather than in a global dashboard, and that explaining per-tenant usage matters more than choosing a particular model. The example code routes requests through an OpenAI-compatible endpoint so the provider details stay isolated behind an adapter. It checks estimated input size against the remaining budget, retries rate-limited requests, and returns a review-needed status when budgets are exceeded or responses fail.

0
ProgrammingDEV Community ·

Claude Code Scheduled Tasks Silently Stop in Three Ways, Study Finds

A developer running Claude Code's scheduled tasks unattended on a desktop app found the jobs stopped in three distinct ways. Runs were skipped when the machine was asleep or the app closed, though they executed on the next launch. Some runs died mid-execution after a dropped API connection, with no retry mechanism, so the slot was lost. Some slots were also silently switched to disabled without anyone intending it. The author notes none of these failures appear in logs of the affected runs, so they judge success by whether output artifacts reached git rather than whether a run occurred.