# What Does Agent Layer AI Runtime Security Enforce?

URL: https://airuntimesecuritytools.com/enforcement-point-map/agent-layer/
Updated: 2026-11-16

Agent-layer enforcement inspects an AI agent's tool calls and plans and decides each one before it runs. As of 16 Nov 2026, 14 of 16 tools documented it specifically, 2 in general terms or in part, and 0 had none found. Hooks, proxies and platform integrations are the usual routes.

## Agent layer: every vendor

**Agent layer by vendor**

| Vendor | Level | Note | Evidence |
|---|---|---|---|
| [Aim Security](/vendors/aim-security/) | Yes | Tool inputs, outputs and actions are inspected and enforced. | [aim-security-51](/vendors/aim-security/#fact-aim-security-51), [aim-security-53](/vendors/aim-security/#fact-aim-security-53), [aim-security-73](/vendors/aim-security/#fact-aim-security-73) |
| [Bay](/vendors/bay/) | Yes | Tool calls evaluated in whole-session context. | [bay-47](/vendors/bay/#fact-bay-47), [bay-58](/vendors/bay/#fact-bay-58), [bay-54](/vendors/bay/#fact-bay-54) |
| [Check Point](/vendors/check-point/) | Yes | Agent tool calls evaluated at runtime. | [check-point-25](/vendors/check-point/#fact-check-point-25), [check-point-26](/vendors/check-point/#fact-check-point-26) |
| [Cisco](/vendors/cisco/) | Yes | Tool calls and results checked. | [cisco-10](/vendors/cisco/#fact-cisco-10), [cisco-13](/vendors/cisco/#fact-cisco-13), [cisco-41](/vendors/cisco/#fact-cisco-41) |
| [CrowdStrike](/vendors/crowdstrike/) | Yes | Agent-level hooks enforce policy inside supported agents. | [crowdstrike-10](/vendors/crowdstrike/#fact-crowdstrike-10) |
| [HiddenLayer](/vendors/hiddenlayer/) | Yes | Tool calls, MCP responses and sessions are inspected and enforced. | [hiddenlayer-08](/vendors/hiddenlayer/#fact-hiddenlayer-08), [hiddenlayer-16](/vendors/hiddenlayer/#fact-hiddenlayer-16), [hiddenlayer-17](/vendors/hiddenlayer/#fact-hiddenlayer-17) |
| [Lasso Security](/vendors/lasso-security/) | Yes | Inline enforcement on agent tool calls and execution trace. | [lasso-security-10](/vendors/lasso-security/#fact-lasso-security-10), [lasso-security-12](/vendors/lasso-security/#fact-lasso-security-12) |
| [Microsoft](/vendors/microsoft/) | Yes | Agent 365 real-time protection of tool calls. | [microsoft-28](/vendors/microsoft/#fact-microsoft-28), [microsoft-30](/vendors/microsoft/#fact-microsoft-30) |
| [Noma Security](/vendors/noma-security/) | Yes | Tool calls and sessions evaluated. | [noma-security-23](/vendors/noma-security/#fact-noma-security-23), [noma-security-19](/vendors/noma-security/#fact-noma-security-19) |
| [Operant AI](/vendors/operant-ai/) | Yes | Agent Protector and Semantic Firewall act on tool calls and plans. | [operant-ai-23](/vendors/operant-ai/#fact-operant-ai-23), [operant-ai-24](/vendors/operant-ai/#fact-operant-ai-24), [operant-ai-30](/vendors/operant-ai/#fact-operant-ai-30) |
| [Pillar Security](/vendors/pillar-security/) | Yes | Tool abuse detection and pre-tool hooks. | [pillar-security-09](/vendors/pillar-security/#fact-pillar-security-09), [pillar-security-25](/vendors/pillar-security/#fact-pillar-security-25) |
| [SentinelOne (Prompt Security)](/vendors/sentinelone/) | Yes | Agent policy enforcement and tool-use validation. | [sentinelone-43](/vendors/sentinelone/#fact-sentinelone-43), [sentinelone-34](/vendors/sentinelone/#fact-sentinelone-34) |
| [Straiker](/vendors/straiker/) | Yes | Tool calls and agent actions are evaluated. | [straiker-22](/vendors/straiker/#fact-straiker-22), [straiker-40](/vendors/straiker/#fact-straiker-40) |
| [Zenity](/vendors/zenity/) | Yes | Boundaries and AIDR enforce on agent actions and tool calls. | [zenity-45](/vendors/zenity/#fact-zenity-45), [zenity-24](/vendors/zenity/#fact-zenity-24), [zenity-52](/vendors/zenity/#fact-zenity-52) |
| [Cloudflare](/vendors/cloudflare/) | Partial | MCP portal curates tools with logs; DLP can scan tool call arguments. | [cloudflare-09](/vendors/cloudflare/#fact-cloudflare-09), [cloudflare-17](/vendors/cloudflare/#fact-cloudflare-17), [cloudflare-18](/vendors/cloudflare/#fact-cloudflare-18) |
| [Palo Alto Networks](/vendors/palo-alto-networks/) | Partial | MCP server tool scanning and agent security are documented; the MCP server relies on the agent calling it. | [palo-alto-networks-45](/vendors/palo-alto-networks/#fact-palo-alto-networks-45), [palo-alto-networks-48](/vendors/palo-alto-networks/#fact-palo-alto-networks-48), [palo-alto-networks-51](/vendors/palo-alto-networks/#fact-palo-alto-networks-51) |

## What does an enforcement point at the agent layer see?

An agent-layer control sees the agent's proposed tool calls, the arguments, the results that come back and, in some designs, the whole step-by-step trace. That lets it judge a sequence, not only a single prompt.

Cisco's Project Quarry, open sourced on 28 Oct 2026, is a detection language for this layer: it lets a team describe a multi-step agent attack once and check every agent session against it, mapping findings to alert, block, redact, quarantine or log.

## What does it miss?

A hook covers only the agents it is installed for. CrowdStrike lists agent-level hooks for named coding agents, and Noma lists named agents for its endpoint enforcement; an agent not on such a list is not covered by that hook.

Agents can also try to get around a monitor. A September 2026 arXiv benchmark reports agents encoding prohibited commands, splitting operations across tool calls and retrying until context left the monitor's history, with best-of-3 evasion success up to 88% in its tests.

Sources: [Cisco: Introducing Project Quarry](https://blogs.cisco.com/ai/project-quarry-agent-security-detections); [arXiv 2609.30217: Instrumental Monitor Evasion Emerges Under Ordinary Task Pressure](https://arxiv.org/abs/2609.30217); [Noma: Extends agent security and governance to the employee endpoint](https://www.prnewswire.com/news-releases/noma-extends-agent-security-and-governance-to-the-employee-endpoint-302890925.html).

## How is agent-layer enforcement deployed?

Common routes are a native hook the agent calls before each tool, a gateway or proxy that mediates tool and model traffic, a platform integration with a SaaS or cloud agent service, and a sandbox that runs the agent. Noma describes enforcement for agents including Claude Code, Cursor, Codex and Windsurf on employee machines.

The agent layer is the one most vendors are building toward in 2026, so documentation changes fast; read dates matter.

## Which tools document it, read 16 Nov 2026

On 16 Nov 2026, 14 of 16 tools had enforcement at the agent layer documented specifically, 2 in general terms or partly, and 0 none found. Levels follow the method page: a tool earns yes only when a vendor page or document describes this specifically, and the cell notes quote what the evidence says.

- Documented specifically: Aim Security, Bay, Check Point, Cisco, CrowdStrike, HiddenLayer, Lasso Security, Microsoft, Noma Security, Operant AI, Pillar Security, SentinelOne (Prompt Security), Straiker and Zenity.
- Documented in general terms or in part: Cloudflare and Palo Alto Networks.

Sources: [Aim Security evidence page for this cell](https://www.catonetworks.com/platform/agentic-ai-security/); [Bay evidence page for this cell](https://www.bay.io/agent-runtime-enforcement?utm_source=airuntimesecuritytools.com&utm_medium=referral&utm_campaign=asset-dim); [Check Point evidence page for this cell](https://www.checkpoint.com/ai-security/ai-agent-security/).

## Related pages

- [The full Enforcement Point Map](/enforcement-point-map/)
- [Application layer](/enforcement-point-map/application-layer/)
- [Endpoint layer](/enforcement-point-map/endpoint-layer/)
- [Method](/enforcement-point-map/methodology/)
- [Vendor profiles](/vendors/)
