PrivacyFlow study reports fewer privacy failures in AI agent workflows

Published Monday, November 2, 2026 · 11:49 PM EST

Researchers report that PrivacyFlow, a defense for AI agents handling personal data, increased the share of completed task sequences without a detected privacy violation by 22-55 percentage points across six models. The preprint introduces a benchmark containing 5873 tasks in 986 task sequences, with changing environments and rules governing which sensitive information agents may share. Before an agent executes a task, PrivacyFlow traces planned disclosures, considers recipients' background knowledge and simulates what outsiders could infer. It then minimizes the plan, blocks it or consults the user. Allowing proactive questions can increase authorized task completion, but shifts decision effort back to the user.

Sources

cs.LG updates on arXiv.org

Published by Tech & Business, a media brand covering technology and business. This story was sourced from cs.LG updates on arXiv.org and reviewed by the T&B editorial agent team.

Back to Newswire

Keep reading

Full wire

AI Products

Nvidia reportedly requires cloud clients to direct 51%+ of services outside China

Nvidia verbally told cloud computing clients to sell 51%+ of their services to non-Chinese companies, the Financial Times reported, citing sources.

Security

Attackers target unpatched WordPress payment flaw that exposes wallet credentials

Attackers are attempting to exploit an unpatched vulnerability in the CryptoPayment Gateway WordPress plugin, Defused Cyber said. CVE-2026-81648 has a CVSS score of 10.0 and affects versions 1.2.1 through 1.2.2, the latest version.

Robotics

Tesla tests two dozen Cybercabs on Detroit-area roads

Tesla is testing two dozen Cybercabs from its Southfield facility on roads in the Detroit area, The Detroit News reported.

AI Science

Preprint reports agent context compression with up to 50.68% fewer attention pairs

Researchers report a method that matched or exceeded an uncompressed language-model agent's accuracy on long-running tasks while using up to 50.68% fewer end-to-end attention pairs, a measure of attention computation.

AI Capital

Trajectory raises $45 million and opens model post-training platform

Trajectory has raised $45 million in Series A funding at a $300 million valuation and is making its model post-training platform publicly available, TechCrunch reported. Sequoia led the round, with participation from Nvidia, Bessemer Venture Partners and Bain Capital Ventures.

Robotics AI

Tutor raises $75M Series B led by a16z

Tutor raised a $75M Series B led by a16z, bringing its total funding to $117M, Forbes reported. The company builds AI warehouse robots and is training a foundation model for robotics, putting the financing alongside its work on both robots and their underlying AI.