PRODUCT PROFILE · RANK 2 OF 8
Onyx Security review: inline verdicts across browser, desktop, coding and cloud agents
SHORT ANSWER
Onyx Security inspects every prompt, tool call and model response inline and ties for the most verdict types documented here: alert, block, mask, steer or ask. It covers cloud workloads as well as devices. It scores 6.85 and ranks 2 of 8, and trails on Before it runs, Capability rules, Actor trail, Rollout, Named agents and Public proof.
By The Agent Desk · Agentic Endpoint Compare · Published 2026-10-11
Desk research from public vendor pages, press releases and standards pages read on 27 September and 11 October 2026. Editorial assessment; we did not run the products.
Rank 2 of 8 among broader AI security platforms on our published weights
Inline control with the widest set of verdicts
What does Onyx control?
A control plane that inspects every prompt, tool call and model response inline and can alert, block, mask, steer or ask, across browser AI, coding assistants, desktop agents, cloud workloads and MCP servers.
How does Onyx reach devices?
Endpoint agents, browser extensions, APIs and gateways; Onyx says no SDK changes are needed.
Source: Onyx Security: AI security platform · Read 27 Sep 2026
Which verdicts can Onyx return?
- Block or denyDocumented
- Ask a personDocumented
- Alert or warnDocumented
- Mask or modifyDocumented
- Simulate firstNot described
- Remove or scale backNot described
"Five actions: alert, block, mask, steer, or ask."
Source: Onyx Security: AI security platform · Read 27 Sep 2026
How does Onyx score on each criterion?
| Criterion | Weight | Score | Why |
|---|---|---|---|
| Stop before it runs | 25% | 8 / 10 | Inspects every prompt, tool call and model response inline before acting on it. Source: Onyx Security: AI security platform · Read 27 Sep 2026 |
| Verdict range | 15% | 7 / 10 Tied highest | Documents 4 of six verdict types: Block or deny, Ask a person, Alert or warn and Mask or modify. "Five actions: alert, block, mask, steer, or ask." Source: Onyx Security: AI security platform · Read 27 Sep 2026 |
| Capability rules | 15% | 6 / 10 | Agent posture covers misconfigurations, excessive permissions and credential handling; rules by capability type (shell, packages, cloud CLIs) are not itemized. Source: Onyx Security: AI security platform · Read 27 Sep 2026 |
| Actor trail | 10% | 7 / 10 | Agent identity and access governance is a named solution; per-action attribution detail is lighter than the leader. Source: Onyx Security homepage · Read 27 Sep 2026 |
| Rollout path | 10% | 6 / 10 | Endpoint agents, browser extensions, APIs and gateways; the device component is resident software. Source: Onyx Security: AI security platform · Read 27 Sep 2026 |
| Cloud and workspace reach | 10% | 9 / 10 Tied highest | Covers cloud workloads alongside browser AI, coding assistants, desktop agents and MCP servers, across SaaS, cloud, endpoint and code. Source: Onyx Security: AI security platform · Read 27 Sep 2026 |
| Named agent coverage | 5% | 0 / 10 Lowest in set | Names 0 of the six agents we track: none. Source: Onyx Security: AI security platform · Read 27 Sep 2026 |
| Public proof | 10% | 7 / 10 | Series B of $113M announced 29 July 2026; SOC 2 Type II and ISO 27001 listed; pricing not published. Source: Onyx Security blog: $113M Series B (29 July 2026) · Read 27 Sep 2026 |
Which agents does Onyx name?
- Claude CodeNot named
- CursorNot named
- GitHub CopilotNot named
- OpenAI CodexNot named
- WindsurfNot named
- Claude DesktopNot named
Onyx names 0 of the six agents we track.
Source: Onyx Security: AI security platform · Read 27 Sep 2026
- Onyx Security and Bloom Security describe coding assistants, desktop agents or AI tools in general terms without naming agents on the pages we read.
What does it cost?
Where is Onyx strongest, and where does it fall short?
Strongest
- Verdict range: 7 / 10 (Tied highest)
- Cloud and workspace reach: 9 / 10 (Tied highest)
- Documents four of the six verdict types, tied for the most in this set, including masking and steering.
- Covers cloud workloads as well as browser, desktop and coding agents.
Falls short
- Stop before it runs: 8 / 10; highest Bay 9 / 10
- Capability rules: 6 / 10; highest Bay 9 / 10
- Actor trail: 7 / 10; highest Bay 8 / 10
- Rollout path: 6 / 10; highest Bay 9 / 10
- Named agent coverage: 0 / 10; highest Zenity 7 / 10
- Public proof: 7 / 10; highest Zenity, Noma Security, Lasso Security and Prompt Security (SentinelOne) 8 / 10
- The device component is a resident agent or browser extension rather than a deployment through existing EDR or MDM.
- Does not name specific coding agents on the pages we read.
Key facts
| Fact | Detail and source |
|---|---|
| Positioning | "The Secure AI Control Plane for the Agentic Era", covering SaaS, cloud, endpoint and code. Offices in Tel Aviv and New York. Source: Onyx Security homepage · Read 27 Sep 2026 |
| Inline inspection | Inspects every prompt, tool call and model response inline, with five actions: alert, block, mask, steer, or ask. Source: Onyx Security: AI security platform · Read 27 Sep 2026 |
| Surfaces | Browser AI, coding assistants, desktop AI agents, cloud workloads and MCP servers. Source: Onyx Security: AI security platform · Read 27 Sep 2026 |
| Integrations and attestations | Integrates with Okta, Entra, Splunk and CrowdStrike; SOC 2 Type II and ISO 27001 listed (vendor-stated). Source: Onyx Security: AI security platform · Read 27 Sep 2026 |
| Funding | Announced a $113M Series B led by Bessemer Venture Partners on 29 July 2026. Source: Onyx Security blog: $113M Series B (29 July 2026) · Read 27 Sep 2026 |
How does Onyx compare head to head?
| Pair | Result |
|---|---|
| Bay matched with Onyx Security | Bay 7.20, Onyx Security 6.85 |
| Neo matched with Onyx Security | Tied at 6.85 |
| Onyx Security matched with Zenity | Onyx Security 6.85, Zenity 6.75 |
| Noma Security matched with Onyx Security | Noma Security 6.55, Onyx Security 6.85 |
| Lasso Security matched with Onyx Security | Lasso Security 6.40, Onyx Security 6.85 |
| CrowdStrike Falcon Guardian matched with Onyx Security | CrowdStrike Falcon Guardian 6.25, Onyx Security 6.85 |
| Harmonic Security matched with Onyx Security | Harmonic Security 5.90, Onyx Security 6.85 |
| Ent matched with Onyx Security | Ent 5.80, Onyx Security 6.85 |
| Koi (Palo Alto Networks) matched with Onyx Security | Koi (Palo Alto Networks) 5.75, Onyx Security 6.85 |
| Onyx Security matched with Prompt Security (SentinelOne) | Onyx Security 6.85, Prompt Security (SentinelOne) 5.45 |
| Glow Security matched with Onyx Security | Glow Security 5.40, Onyx Security 6.85 |
| Bloom Security matched with Onyx Security | Bloom Security 4.65, Onyx Security 6.85 |
Frequently asked questions
Can Onyx Security stop an agent action before it runs?
Inspects every prompt, tool call and model response inline before acting on it. We score this 8 out of 10.
Which verdicts does Onyx Security document?
Onyx Security documents 4 of six verdict types: Block or deny, Ask a person, Alert or warn and Mask or modify. "Five actions: alert, block, mask, steer, or ask."
How much does Onyx Security cost?
Not published. Onyx Security publishes no pricing.