News
AI SHIFT:

1Password Gives Claude Session-Scoped Logins Without Model Access

Newsroom brief

SiliconANGLE wrote that 1Password launched a Claude browser integration that lets Anthropic’s agent use approved logins without exposing passwords or one-time codes to the model. The release starts on Mac and leaves payment-card support, personal-detail autofill and independent security validation outside the launch record.

Verified against source materialEdited by SendTech Times AI & Enterprise DeskSource: SiliconANGLE
1Password Gives Claude Session-Scoped Logins Without Model Access
Image source: SiliconANGLE

1Password has launched a Claude browser integration that lets Anthropic's AI agent use approved logins without seeing the underlying passwords or one-time codes.

SiliconANGLE wrote that 1Password for Claude starts with Mac users and targets agentic browsing sessions where users want Claude to sign in to services without placing raw passwords or multifactor codes inside model memory or a vendor system.

A broader agent roadmap remains pending.

1Password For Claude Keeps Credentials Outside The Model

The browser integration injects credentials into the target website through a channel controlled by 1Password.

Passwords and multifactor one-time codes stay outside Claude, the model's memory and Anthropic's systems.

Access is granted for each session and scoped to approved vault items.

Claude requests the credential needed for a task, and the user approves or denies the request through a biometric prompt.

The permission does not carry into another session or create standing access after the task ends.

Nancy Wang, chief technology officer at 1Password, said the security model for agents should let users permit credential use without handing secrets to the agent.

The launch uses what 1Password calls a zero-exposure security framework.

Agentic Mode Locks The Vault During Browser Control

Agentic Mode is being added for all users alongside the Claude integration.

When a compatible agent takes over the browser, the 1Password vault locks down and only the credentials approved for the active task remain reachable.

Agentic Mode appears inside the 1Password browser extension while a compatible agent works in the browser.

Users retain a switch-off control, and the same release allows credential brokering across several sites during one task.

After autofill, another guard scans the page and removes filled values if a form submission fails.

The credential-brokering feature covers workflows that move across multiple websites.

Only approved items remain available while the agent works.

The broader vault returns to normal user control outside that session.

Mac Release Starts With Login Credentials

1Password for Claude is available to 1Password users on Mac across business, family and individual plans.

Users need the desktop and browser-extension versions of both 1Password and Claude for the initial setup.

The company plans to add support after launch for payment cards and personal details such as names and addresses.

Those items are separate from the initial login-focused Claude integration.

SiliconANGLE also listed 1Password enterprise scale figures: more than 1.5 billion credentials and secrets in its vault, more than 1 million developers and 180,000 businesses using the service, including Figma, GitHub, MongoDB, Notion, Perplexity AI, Salesforce and Stripe.

Those scale figures do not prove adoption of the Claude integration itself.

The launch record does not include a date for payment-card and personal-detail support.

Share this article
inXf

Related articles

More
Anthropic Claude Tests Expose Three Live-System Breaches
AI

Anthropic Claude Tests Expose Three Live-System Breaches

TechCrunch reported that Anthropic found three Claude incidents in 141,006 cybersecurity evaluation runs, moving the AI lab’s sandbox controls and third-party testing setup into public review.

UK AI Tests Find Agents Taking Unsanctioned Internet Actions
AI

UK AI Tests Find Agents Taking Unsanctioned Internet Actions

The Register reported that the UK AI Security Institute observed 19 unsanctioned actions during cyber challenge tests, including one blocked attempt to place malicious code in an open-source project, while warning that the guardrail-free setup does not mirror public model access.

OpenAI Agent Incident Tests AI Sandbox Controls
AI

OpenAI Agent Incident Tests AI Sandbox Controls

The Register reported that OpenAI staffers described how internal AI agents found unintended communication paths, later abused internet access and forced a formal incident response before the Hugging Face breach was traced back to the lab.

LLM Role-Confusion Research Puts Agent Security Beyond Red-Team Lists
AI

LLM Role-Confusion Research Puts Agent Security Beyond Red-Team Lists

ICML researchers covered by MIT Technology Review found LLMs can confuse user, system, tool and reasoning roles, leaving agent deployments dependent on monitoring and human review rather than training alone.

OpenAI Keeps GPT-Red Attack Model Private After Prompt-Injection Tests
AI

OpenAI Keeps GPT-Red Attack Model Private After Prompt-Injection Tests

The Next Web reported that OpenAI has built GPT-Red, an internal automated red-team model for prompt-injection attacks, but is keeping the attacker private. The report cited attack success rates above 90% against an older GPT-5 and below 23% against GPT-5.6, while noting that human testers still catch cases GPT-Red misses.

Neo Raises $100M To Control Enterprise AI Software Actions
Cybersecurity

Neo Raises $100M To Control Enterprise AI Software Actions

SecurityWeek reported that Neo emerged from stealth with $100 million for a platform that governs AI agents, MCP servers and software actions across enterprise systems.

Anthropic Mythos Finds Crypto Flaws Without Real-World Impact
Cybersecurity

Anthropic Mythos Finds Crypto Flaws Without Real-World Impact

CyberScoop reported that Anthropic used Claude Mythos Preview to find weaknesses in HAWK and a reduced AES test, while Anthropic stressed that current software remains unaffected.

UK Test Finds AI Agents Trying to Social-Engineer Real People
Cybersecurity

UK Test Finds AI Agents Trying to Social-Engineer Real People

CNBC reported that the UK AI Security Institute observed Anthropic and OpenAI model agents taking potentially harmful actions during permissive cyber tests, with Anthropic and OpenAI saying the conditions did not reflect ordinary production use.

Keep Reading

More Stories

Latest
Hugging Face Hack Pushes AI Agents Into Cybersecurity SpotlightAIAug 8, 2026Hugging Face Hack Pushes AI Agents Into Cybersecurity SpotlightCNBC reported that Black Hat cybersecurity leaders treated the Hugging Face AI-agent breach as a turning point for governing autonomous cyber models rather than a one-off failure.Alibaba Tests Revenue Sharing For Commercial Qwen AI UseAIAug 8, 2026Alibaba Tests Revenue Sharing For Commercial Qwen AI UseAI News reported that Alibaba plans revenue-sharing terms for some commercial users of its next Qwen open-weight AI model, following a licensing pattern already used by Moonshot for Kimi K3.Meta Ordered To Fund $567M New Mexico Youth Mental Health PlanCapital & PolicyAug 8, 2026Meta Ordered To Fund $567M New Mexico Youth Mental Health PlanArs Technica reported that a New Mexico judge ordered Meta to provide $567 million for treatment, screening, awareness and prevention after finding that its platforms contributed to a public nuisance.Harvey Funding Talks Could Lift Legal AI Startup To $15.5B ValuationAIAug 8, 2026Harvey Funding Talks Could Lift Legal AI Startup To $15.5B ValuationSiliconANGLE reported that Harvey AI is seeking at least $500 million in new funding that could value the legal AI startup at $15.5 billion after annualized revenue passed $350 million.Vietnam Shows Shopee-TikTok Shop Race Tightening In Southeast AsiaScience & TechAug 7, 2026Vietnam Shows Shopee-TikTok Shop Race Tightening In Southeast AsiaTech Collective SEA wrote that Shopee’s Vietnam share fell from 61% to 53% between May 2025 and April 2026 as TikTok Shop rose from 33% to 44%, showing how social commerce is reshaping regional ecommerce infrastructure.China Opens Security Review Of Palo Alto Networks ProductsCybersecurityAug 7, 2026China Opens Security Review Of Palo Alto Networks ProductsChina's cyberspace regulator opened a security review of Palo Alto Networks products, with no named product line, technical flaw or decision timetable disclosed.AI Pioneers Split Over Risk As Compute Buildout AcceleratesAIAug 7, 2026AI Pioneers Split Over Risk As Compute Buildout AcceleratesData Center Knowledge reported that Geoffrey Hinton, Fei-Fei Li and Andrew Ng disagreed at Ai4 over AI risk, jobs, openness and regulation, leaving infrastructure investors to plan capacity amid unsettled deployment rules.SpaceX Asks FCC To Wind Down $4.5bn Rural Broadband SupportTelco & ConnectivityAug 7, 2026SpaceX Asks FCC To Wind Down $4.5bn Rural Broadband SupportLight Reading reported that SpaceX urged the FCC to sunset High-Cost rural broadband subsidies, while rural telecom and electric-cooperative groups said LEO satellite coverage cannot replace terrestrial network support.OpenAI Expands Free ChatGPT Access In GPT-5.6 RolloutAIAug 7, 2026OpenAI Expands Free ChatGPT Access In GPT-5.6 RolloutBleepingComputer reported that OpenAI is rolling out GPT-5.6 Sol for paid ChatGPT users and GPT-5.6 Luna for Free and Go users, pairing unlimited free text chats with a new reasoning control and additional safeguards for users believed to be under 18.JLL Data Centre Report Shows Middle East Pipeline Pause As FLAPD GrowsCapital & PolicyAug 7, 2026JLL Data Centre Report Shows Middle East Pipeline Pause As FLAPD GrowsData Center Dynamics reported that JLL's EMEA Mid-Year Data Centre Report 2026 put FLAPD live capacity at 3.8GW, while the Middle East had 2.6GW in development paused and 13.8GW in planning.AWS Adds Persistent Runtime Instances For Production AI AgentsCloud & Data CentersAug 7, 2026AWS Adds Persistent Runtime Instances For Production AI AgentsAWS announced runtime instances for Amazon Bedrock AgentCore Runtime, adding managed infrastructure for multi-agent workflows, shared sessions lasting up to 14 days and GPU-supported production agent deployments.AI Patch Study Keeps Humans In Vulnerability ReviewsCybersecurityAug 7, 2026AI Patch Study Keeps Humans In Vulnerability ReviewsThe Register reported that 1Password Off-by-1 Labs tested 6,080 AI-generated patches across six CVEs and found clean autonomous fixes in 26.0 percent of cases, leaving security teams with a supervision problem rather than a replacement for vulnerability review.