/
← Voltar ao feed
low

Proofpoint Protect 2026: AI Agent Security Must Shift to Intent-Based Access Control and Detection

1 fonte(s) · 1 atualização(ões) · publicado 30/09/2026 13:11 · coletado há 4 dias

At Proofpoint Protect 2026, company executives and researchers argued that AI is raising the speed and scale of attacks but not rewriting the attacker playbook. They said security for AI agents must move beyond identity and behavior checks to evaluating intent. Proofpoint presented its Agentic Data & AI Security System, Semantic Business Policies, Intent-Based Access Control and the Nexus Intent-Based Detection Model as its response. Researchers also flagged TeamFiltration abuse and expect more indirect prompt injection attacks in the wild. The article is conference coverage and vendor positioning, not a specific incident.

Detalhes técnicos

  • Threat researchers Daniel Blackford and Sarah Sabotka said attackers use AI to improve phishing, make translated lures more convincing and reach more regions. Operational security can still suffer, and AI can make attackers sloppy.
  • Yaniv Miron said one TeamFiltration scenario could reach an organization's crown jewels in about 90 seconds without triggering account lockouts.
  • Miron described an indirect prompt-injection attack that delivered its payload through a Google Calendar invitation and later modified the invitation to remove traces. He called most current activity experimentation but expects more in-the-wild attacks in the coming months.
  • Compromised supplier mailboxes can be used to reply inside existing threads with no malicious attachment, suspicious domain or obvious behavioral anomaly. Proofpoint's Nexus model instead evaluates what a message is trying to achieve and whether that makes sense in the relationship context.
  • Ryan Kalember said much of the AI risk Proofpoint sees is not traditional cyber risk. It includes fraud, prohibited AI use and employee behavior that creates compliance issues.

Remediação

  • Gain visibility into the AI systems and agents in the environment, what they can access and which business processes depend on them.
  • Define business policies that are enforced at runtime for both employees and agents. Examples include requiring approval for refunds even when an agent technically has permission.
  • Evaluate agents against human baselines (time, false positives, false negatives) before handing them security responsibilities, with human-in-the-loop approval for consequential decisions.

Linha do tempo

  1. proofpoint.com ↗ · pub. 30/09/2026 13:11 · coletado 19/10 18:02 primeiro registro

    Proofpoint Protect 2026: AI Agent Security Must Shift to Intent-Based Access Control and Detection

    ler a análise desta fonte

    At Proofpoint Protect 2026, company executives and researchers argued that AI is raising the speed and scale of attacks but not rewriting the attacker playbook. They said security for AI agents must move beyond identity and behavior checks to evaluating intent. Proofpoint presented its Agentic Data & AI Security System, Semantic Business Policies, Intent-Based Access Control and the Nexus Intent-Based Detection Model as its response. Researchers also flagged TeamFiltration abuse and expect more indirect prompt injection attacks in the wild. The article is conference coverage and vendor positioning, not a specific incident.

    Technical Details

    • Threat researchers Daniel Blackford and Sarah Sabotka said attackers use AI to improve phishing, make translated lures more convincing and reach more regions. Operational security can still suffer, and AI can make attackers sloppy.
    • Yaniv Miron said one TeamFiltration scenario could reach an organization's crown jewels in about 90 seconds without triggering account lockouts.
    • Miron described an indirect prompt-injection attack that delivered its payload through a Google Calendar invitation and later modified the invitation to remove traces. He called most current activity experimentation but expects more in-the-wild attacks in the coming months.
    • Compromised supplier mailboxes can be used to reply inside existing threads with no malicious attachment, suspicious domain or obvious behavioral anomaly. Proofpoint's Nexus model instead evaluates what a message is trying to achieve and whether that makes sense in the relationship context.
    • Ryan Kalember said much of the AI risk Proofpoint sees is not traditional cyber risk. It includes fraud, prohibited AI use and employee behavior that creates compliance issues.

    Remediation

    • Gain visibility into the AI systems and agents in the environment, what they can access and which business processes depend on them.
    • Define business policies that are enforced at runtime for both employees and agents. Examples include requiring approval for refunds even when an agent technically has permission.
    • Evaluate agents against human baselines (time, false positives, false negatives) before handing them security responsibilities, with human-in-the-loop approval for consequential decisions.