/
← Voltar ao feed
low

Gen Digital sponsored piece promotes Agent Trust Hub amid AI agent prompt injection and malicious skill risks

1 fonte(s) · 1 atualização(ões) · publicado 23/10/2026 11:00 · coletado há 3 horas

This sponsored article by Gen Digital argues that AI agents with delegated permissions are exposed to manipulation through untrusted content. It cites Gen Threat Labs research in which an agent reading email was steered by prompt injection into running system commands, sending data out of its environment and performing destructive file operations, with no software vulnerability involved. It also references Warden Stealer, documented in October, which targets configuration files and local data of AI assistants and coding tools, potentially exposing credentials, access tokens and conversation histories. Gen further reports that 12% of skills on ClawHub were malicious at the time of its earlier research. The bulk of the piece promotes Gen's Agent Trust Hub and related products rather than reporting a new incident.

Detalhes técnicos

  • Prompt injection via email content led an agent to execute commands, exfiltrate data and delete files (controlled experiment).
  • Malicious ClawHub skills impersonated legitimate extensions and carried instructions meant to trick agents into downloading malware.
  • Warden Stealer targets AI assistant and coding tool configuration files and local data.
  • Gen's product set: skill scanning and a curated marketplace, VPN for Agents (Norton VPN infrastructure), Sage (open-source runtime hook layer for Claude Code, Cursor, VS Code, OpenClaw and OpenCode), Norton AI Agent Protection, the AI Data Checker, and the AARTS open specification.
  • Gen collaborated with AMTSO on guidelines for testing agentic-system protection, published in September 2026.

Remediação

  • Scan skills and plugins before installation and prefer verified or curated sources.
  • Enforce runtime checks on shell commands, file operations and web requests before execution.
  • Limit what agents can access and treat content such as email as untrusted input.
  • Layered controls are needed; the article notes no single layer catches every risk and Sage does not replace endpoint protection.

Linha do tempo

  1. bleepingcomputer.com ↗ · pub. 23/10/2026 11:00 · coletado 23/10 12:00 primeiro registro

    Gen Digital sponsored piece promotes Agent Trust Hub amid AI agent prompt injection and malicious skill risks

    ler a análise desta fonte

    This sponsored article by Gen Digital argues that AI agents with delegated permissions are exposed to manipulation through untrusted content. It cites Gen Threat Labs research in which an agent reading email was steered by prompt injection into running system commands, sending data out of its environment and performing destructive file operations, with no software vulnerability involved. It also references Warden Stealer, documented in October, which targets configuration files and local data of AI assistants and coding tools, potentially exposing credentials, access tokens and conversation histories. Gen further reports that 12% of skills on ClawHub were malicious at the time of its earlier research. The bulk of the piece promotes Gen's Agent Trust Hub and related products rather than reporting a new incident.

    Technical Details

    • Prompt injection via email content led an agent to execute commands, exfiltrate data and delete files (controlled experiment).
    • Malicious ClawHub skills impersonated legitimate extensions and carried instructions meant to trick agents into downloading malware.
    • Warden Stealer targets AI assistant and coding tool configuration files and local data.
    • Gen's product set: skill scanning and a curated marketplace, VPN for Agents (Norton VPN infrastructure), Sage (open-source runtime hook layer for Claude Code, Cursor, VS Code, OpenClaw and OpenCode), Norton AI Agent Protection, the AI Data Checker, and the AARTS open specification.
    • Gen collaborated with AMTSO on guidelines for testing agentic-system protection, published in September 2026.

    Remediation

    • Scan skills and plugins before installation and prefer verified or curated sources.
    • Enforce runtime checks on shell commands, file operations and web requests before execution.
    • Limit what agents can access and treat content such as email as untrusted input.
    • Layered controls are needed; the article notes no single layer catches every risk and Sage does not replace endpoint protection.