Runtime block as an AI agent control point
Runtime block is the control that stops an agent action before it completes. On 16 Nov 2026, 16 of 16 vendors had a yes for this control point, 0 had partial and 0 had none found. Levels are from vendor documentation read that day, not lab tests.
Facts read on .
Runtime block: every vendor
| Vendor | Level | Note | Evidence |
|---|---|---|---|
| Aim Security | Yes | Risky tool calls and actions are blocked before they execute. | aim-security-51, aim-security-52, aim-security-41 |
| Bay | Yes | Allow, ask, deny or steer before an action completes. | bay-47, bay-48, bay-55 |
| Check Point | Yes | Tool calls evaluated and blocked at runtime via the Guard API. | check-point-25, check-point-26 |
| Cisco | Yes | Blocks prompts, shell commands and tool calls before they run. | cisco-10, cisco-05, cisco-02 |
| CrowdStrike | Yes | Hooks block malicious activity before it runs; Kill Switch blocks model access at gateways. | crowdstrike-10, crowdstrike-11, crowdstrike-24 |
| Island | Yes | Inline control points can stop an action before it happens. | island-16, island-17 |
| Lasso Security | Yes | Says it blocks or quarantines actions before they execute. | lasso-security-10, lasso-security-11, lasso-security-12 |
| Microsoft | Yes | Custom rules block tool calls before execution; endpoint runtime block in preview. | microsoft-28, microsoft-19 |
| Noma Security | Yes | Blocks actions per policy; options depend on integration (a20). | noma-security-20, noma-security-23 |
| Operant AI | Yes | Inline blocking of out-of-scope and harmful actions before execution. | operant-ai-23, operant-ai-24, operant-ai-30 |
| Palo Alto Networks | Yes | Block actions stop requests; SaaS agents can be paused or terminated. | palo-alto-networks-34, palo-alto-networks-47, palo-alto-networks-48 |
| Pillar Security | Yes | Pre-tool hook blocking for Claude Code, Codex CLI and Cursor. | pillar-security-09, pillar-security-10 |
| SentinelOne (Prompt Security) | Yes | States it stops violations before they propagate and enforces at the point of interaction. | sentinelone-43, sentinelone-41 |
| Straiker | Yes | Blocks named actions at supported execution points. | straiker-22, straiker-23 |
| WitnessAI | Yes | Denies off-list tool calls before they execute and blocks prompt injection before agents act. | witnessai-14, witnessai-35 |
| Zenity | Yes | Boundaries and AIDR block actions before they complete. | zenity-45, zenity-46, zenity-52 |
What is runtime block as a control point?
A runtime block evaluates a proposed action, such as a tool call, a file write or an outbound request, and stops it before it completes. It differs from alerting, which reports an action after it happened. It needs a hook where the action can be intercepted.
Sources: Protect enterprise coding agents with Cisco DefenseClaw and NVIDIA OpenShell.
Why it matters and where it stops
Cisco says DefenseClaw checks every prompt, tool call and tool result inside an OpenShell sandbox and decides each web destination by host name. Operant says its Agent Protector can redact a field from a tool result, block a single outbound call or stop one write while the rest of the task completes.
A block only covers the agents and paths that carry the hook. Operant says its Purview path for Copilot detects and alerts but does not sit inline, so the same vendor can have both a blocking and a detect-only route.
Sources: Protect enterprise coding agents with Cisco DefenseClaw and NVIDIA OpenShell; Operant: coverage for Microsoft Copilot and embedded AI assistants.
Who documents it
Yes on 16 Nov 2026: Aim Security, Bay, Check Point, Cisco, CrowdStrike, Island, Lasso Security, Microsoft, Noma Security, Operant AI, Palo Alto Networks, Pillar Security, SentinelOne (Prompt Security), Straiker, WitnessAI, Zenity.
Partial: none.
None found: none.