Skip to content
Cut to the source.The real-time AI news wire.
Saved in a cookie.

Search tips

  • "exact phrase" the words together, in this order
  • -word leave out stories with the word
  • topic:robotics a topic by ID or name; quote names with spaces: topic:"open-weight models"
  • category:"AI agents" a category
  • company:openai a company, person, product or organization (also entity:)
  • source:forbes.com a publisher by site or name
  • type:analysis news, analysis, opinion, research paper and other kinds
  • after:2026-11-01 stories on or after that New York day; before:2026-11-10 on or before it
Subscribe
NewsletterAboutModels & ResearchAI AgentsAI DevelopmentInfrastructure & HardwareImage, Video & AudioRobotics & Autonomous SystemsBusiness & IndustryPolicy & RegulationSafety & SecurityBenchmarks & EvaluationOpen Source & CommunityAI in Society
Models & ResearchAI AgentsAI DevelopmentInfrastructure & HardwareImage, Video & AudioRobotics & Autonomous SystemsBusiness & IndustryPolicy & RegulationSafety & SecurityBenchmarks & EvaluationOpen Source & CommunityAI in Society
Robotics & Autonomous SystemsBusiness & IndustryPolicy & RegulationSafety & SecurityBenchmarks & EvaluationOpen Source & CommunityAI in Society

14 NovSafety & Securitydevotd.net

Microsoft Research Lazarus: Dormant neuron backdoors, agent permissions

On a white background, blue-gray circles and one teal circle on the left are linked by navy lines. An elbow leader from the 'Dormant neuron' label above points to the teal circle. A teal dashed arrow leads from this circle to a navy robot in the center. A solid navy arrow leads from the robot to a blue-gray padlock on the right. Above the padlock is an 'Agent permissions' label with an elbow leader.

AIAI-generated summary Microsoft researchers introduced Lazarus, a technique that alters specific neurons to plant hidden malicious behavior in AI agents. The company reported attack success rates of 98.80% and 99.70% for a variant tested on Nemotron-Terminal-8B.

Microsoft Agent Frameworks Model Security

Earlier:
Microsoft Researchers Show AI Agents Can Be Backdoored by…
Read at devotd.net

Share onXLinkedInBlueskyEmail

Latest stories

  • Databricks Genie Targets Enterprise Context for AI Agents

    20 November 2026

    aitop.news — At DATA + AI LIVE, OpenAI Co-founder and President Greg Brockman joined Databricks Co-founder and CEO Ali Ghodsi to discuss why even the most capable AI models…

  • Versine | App

    20 November 2026

    creatorstoolbox.com — A fintech app that lets ChatGPT, Claude, Muse, Instinct, or any AI you use send and receive money within limits you set, with every payment approved by you.

  • Google Names 10 African AI Startups for Lab

    20 November 2026

    riotimesonline.com — Google picked 10 startups from thousands in 45+ African countries for its Applied AI Lab in Accra. See who made the cohort and what is still unknown.

  • Web Search APIs for AI Agents Compared: Six Options Ranked

    20 November 2026

    web3aiblog.com — Six web search APIs for AI agents scored on cost per 1,000 searches, index, extraction and privacy, using prices and changelogs checked on 18 November 2026.

  • Buy a Custom Domain in Google AI Studio (2026 Guide)

    20 November 2026

    explainx.ai — Google AI Studio adds in-app custom domain purchase on paid projects. Pricing, steps, Starter Tier limits, and how it compares to July free subdomains.

TruncatedTruncated

Continuous coverage across AI research, agentic systems, infrastructure, and regulation.

Content

  • Latest
  • Archive
  • Topics
  • Newsletter
  • RSS Feed

About

  • About
  • Developers
  • Publishers
  • Advertise

Legal

  • Impressum
  • Privacy Policy
  • Terms & Conditions
  • Disclaimer
  • Contact

© 2026 Truncated. All rights reserved.

Truncated™ is a trademark of Three Things Media.