/
← Voltar ao feed
low

Check Point promotes custom plain-language guardrails for AI apps and agents as enforceable business-rule and DLP controls

1 fonte(s) · 1 atualização(ões) · publicado 22/10/2026 10:00 · coletado há 19 horas

Check Point published a vendor blog post promoting custom guardrails, a feature of Check Point AI Security that lets organizations write business-specific AI rules in plain language, test them and enforce them on every message and tool call. It cites Check Point's 2026 Cloud Security Report, which found that only 14 percent of organizations have formal GenAI policies that are actively enforced and audited. The post uses non-attacker incidents to argue that costly chatbot failures break rules no vendor ships: Air Canada was held liable for a refund its chatbot promised, New York City's small-business assistant gave unlawful tip advice, and a DPD chatbot swore at a customer. It is product and marketing content with no vulnerability, campaign or active threat.

Detalhes técnicos

  • The builder's AI assistant compiles a plain-language description into a policy of meaning-based rules (three sensitivity levels, multilingual), pattern rules for exact formats and combining logic.
  • Once saved, guardrails run without an LLM in the loop and make no external calls. Check Point states a latency of about 35 to 40 ms at the 95th percentile for a policy with one or two rules.
  • Rules can be scoped to user messages, model replies, or the inputs and outputs of agent tool calls. Examples include national ID DLP checks on tickets and outbound email, and a database-drop pattern applied only to a database tool.
  • Self-hosted deployment is stated to be available from November 2026.
  • Stated limitations: one message at a time, so no cross-conversation context; not intended for adversarial threats (prompt attack detectors cover those); judging generated code safety remains hard.
  • The post argues that system-prompt instructions, keyword/regex lists and second-model judging each fall short on enforceability, meaning or latency.

Remediação

  • Run new custom guardrails in detect mode first and review what they flag before switching to blocking.
  • Pair custom guardrails with dedicated prompt-attack detectors for adversarial threats.
  • Define jurisdiction-specific DLP categories (for example Spanish DNI/NIE, Korean resident registration numbers, US health identifiers) rather than relying on US-centric lists.

Linha do tempo

  1. blog.checkpoint.com ↗ · pub. 22/10/2026 10:00 · coletado 22/10 11:00 primeiro registro

    Check Point promotes custom plain-language guardrails for AI apps and agents as enforceable business-rule and DLP controls

    ler a análise desta fonte

    Check Point published a vendor blog post promoting custom guardrails, a feature of Check Point AI Security that lets organizations write business-specific AI rules in plain language, test them and enforce them on every message and tool call. It cites Check Point's 2026 Cloud Security Report, which found that only 14 percent of organizations have formal GenAI policies that are actively enforced and audited. The post uses non-attacker incidents to argue that costly chatbot failures break rules no vendor ships: Air Canada was held liable for a refund its chatbot promised, New York City's small-business assistant gave unlawful tip advice, and a DPD chatbot swore at a customer. It is product and marketing content with no vulnerability, campaign or active threat.

    Technical Details

    • The builder's AI assistant compiles a plain-language description into a policy of meaning-based rules (three sensitivity levels, multilingual), pattern rules for exact formats and combining logic.
    • Once saved, guardrails run without an LLM in the loop and make no external calls. Check Point states a latency of about 35 to 40 ms at the 95th percentile for a policy with one or two rules.
    • Rules can be scoped to user messages, model replies, or the inputs and outputs of agent tool calls. Examples include national ID DLP checks on tickets and outbound email, and a database-drop pattern applied only to a database tool.
    • Self-hosted deployment is stated to be available from November 2026.
    • Stated limitations: one message at a time, so no cross-conversation context; not intended for adversarial threats (prompt attack detectors cover those); judging generated code safety remains hard.
    • The post argues that system-prompt instructions, keyword/regex lists and second-model judging each fall short on enforceability, meaning or latency.

    Remediation

    • Run new custom guardrails in detect mode first and review what they flag before switching to blocking.
    • Pair custom guardrails with dedicated prompt-attack detectors for adversarial threats.
    • Define jurisdiction-specific DLP categories (for example Spanish DNI/NIE, Korean resident registration numbers, US health identifiers) rather than relying on US-centric lists.