AI Governance Two Ways: Obsidian Security vs dope.security

Two very different things happen in your company under the heading of AI. Your engineering team wires an agent into Bedrock and gives it access to a production data store. Your head of sales pastes the renewal pipeline into personal ChatGPT because the enterprise account was slow to load. Obsidian Security AI governance is built for the first one. dope.security is built for the second. Both are real problems, and they have almost nothing in common.

Both companies have an AI story. They're solving different halves

Obsidian Security governs the AI your company builds and connects. dope.security governs what a human on a laptop does with AI. If you read the two websites back to back and came away confused, that's why.

Obsidian publishes AI Agent Security, Shadow AI, MCP Security, AI Prompt Security, GenAI Data Leakage, AI Threat Detection, Agent Visibility, Agent Governance, Agent Runtime Security, and AI-SPM. The AI platforms they name include Amazon Bedrock, Microsoft Foundry, Anthropic Claude, n8n, Google Vertex AI, OpenAI, Microsoft Copilot, and Salesforce Agentforce. They've also announced an integration with Anthropic's Claude Compliance API. That's governance of agents, integrations, and the identities those agents carry.

dope.security starts from a different question: what did the person at the keyboard just send, and can we stop it before it leaves the device? That's an inline problem, and it needs an inline answer.

Obsidian's lane: the agents and integrations you build

Give Obsidian its due here, because this is genuinely hard work. An AI agent with an OAuth grant is an identity with permissions, running unattended, at machine speed. It doesn't get phished, but it does get over-permissioned, and it does get abandoned in a tenant nobody audits.

Obsidian's approach is API integrations and browser telemetry correlated in a Knowledge Graph that maps identity and activity across SaaS apps, browsers, and identity providers. They describe Data Depth, an Obsidian AI Assistant, self-learning detection models, and Network Effects. Their published scale figures are over 1.5 billion daily events processed, over 59 million unique identities mapped, and over 3.4 thousand monthly threats stopped.

If your AI risk is an agent runtime you own, an MCP server your platform team stood up last quarter, or an Agentforce deployment with reach into customer records, that's Obsidian's lane and we're not going to pretend otherwise.

dope.security's lane: what a human does with AI

Our AI governance runs on a three-layer model, and the layers stack.

Layer one is Shadow IT discovery. You can't govern what you can't see. Before any policy gets written, you need the list of AI tools actually in use, not the list in the procurement spreadsheet.

Layer two is dope.SWG policy. The Fly Direct Secure Web Gateway does SSL inspection and URL filtering on-device, so an AI category or a specific app can be allowed, blocked, or inspected. Policy changes push in seconds, which matters when a new AI tool goes viral internally on a Tuesday afternoon.

Layer three is Cloud Application Control. This is the one people underestimate. CAC restricts access to approved SaaS tenants only, so an employee can sign into your enterprise ChatGPT, Claude, Google, or Microsoft 365 account and cannot sign into their personal one. You keep the productivity. You lose the shadow account. Blocking ChatGPT outright just moves the behavior to a phone, and everybody knows it.

Dopamine DLP: inspecting the prompt and the upload

Layers one through three decide where people can go. Dopamine DLP decides what they can send once they get there.

It's endpoint DLP built into the on-device proxy, and it watches file uploads and AI prompts in real time. An LLM in dopecloud reads what's in the prompt and answers back in about a second, so there's nothing to tune and no rule language to learn. Coverage includes ChatGPT, Claude, Perplexity, Abacus, and Copilot. It's covered by US Patent no. 12,464,023.

Modes are Block, Monitor, and Off, with Warning mode coming soon rather than shipped today. You get per-policy DLP, exceptions for users and groups, bypass lists including dope-managed bypasses, and automatic exemption for tenants already under Cloud Application Control so you're not inspecting traffic you've already sanctioned. That last detail matters more than it sounds: the sanctioned workflow your users actually want stays fast, which is what keeps them on it.

Every violation carries a Dopamine explanation, a plain-language AI summary of why it fired. Violations forward to your SIEM. On privacy, classification runs on OpenAI zero data retention APIs under HIPAA and BAA, so nothing trains on your data and nothing is retained.

AI Usage Analytics: the number your board will ask for

There's an Analytics section in dope.console dedicated to AI. It shows Top AI Applications by transaction volume and by user count, Top AI Users by AI transaction volume and distinct apps used, and an Applications-per-User breakdown. Headline totals cover Total AI Requests, Active AI Users, and Distinct AI Apps Detected, all over a rolling 7-day window. You can export a branded PDF on demand.

One honest note: sanctioned versus unsanctioned classification on that view, and policy enforcement directly from it, are roadmap items. They're not shipped. When someone asks whether they can click a row and enforce, the answer today is that you write the policy in dope.SWG or CAC.

The thick-client problem nobody wants to talk about

Here's where architecture decides the outcome.

A growing share of AI usage never touches a browser tab. ChatGPT Desktop and Claude Desktop are native apps. IDEs call AI APIs directly. CLI tools pipe a codebase into a model from a terminal window. Obsidian's picture is assembled from connected applications and telemetry, which means it's strongest where an app is connected, and a terminal window isn't an app you connect.

dope.endpoint takes the other position. Inspection happens on the device rather than inside a browser, with SSL inspection and break/inspect running locally in under 100 MB of RAM, native on Mac and Windows. A native client's HTTPS traffic is still HTTPS traffic leaving the same laptop, so in principle it passes through the same on-device proxy as traffic from a tab. That makes the architecture structurally suited to thick clients rather than dependent on where the user happens to be typing.

The honest caveat: certificate pinning in some native applications will break inspection and require a bypass. That's not a footnote we hide. It's exactly why dope.SWG ships an SSL error notification feature in dope.console, which surfaces the broken traffic so an admin can build the bypass in a few clicks instead of finding out weeks later when a user complains. Ask us to show you that screen in a demo. Ask everyone else what their equivalent is.

Mapping AI risks to the layer that catches them

The bottom third of that table is the honest part. We don't govern agent runtimes on Bedrock or Vertex, and we don't claim MCP security. Our AI-Powered SSPM is scoped to Microsoft 365 and Google tenants, where it discovers third-party OAuth-connected apps and scores them across application metadata, usage telemetry, external vendor research, and tenant-level intelligence, then returns two prioritized actions per app and a one-sentence Dopamine insight.

Where Obsidian Security AI governance is a good fit

If you're a company that builds with AI, Obsidian is a serious answer. Agent visibility, agent governance, agent runtime security, and AI-SPM across Bedrock, Microsoft Foundry, Vertex, OpenAI, Copilot, and Agentforce is a real product surface, and their connector breadth across 200+ enterprise applications is depth we don't claim. Add MCP security and prompt security across those platforms and you've got coverage of a category most SSE vendors, including us, simply don't operate in.

The honest split

If the risk is your agents, that's Obsidian's lane. Over-permissioned service principals, unmonitored MCP servers, agents with standing access to production data, runtime behavior nobody is watching.

If the risk is your employees, that's ours. The prompt with a customer list in it. The upload to a tool procurement has never heard of. The personal ChatGPT login on a managed MacBook. The AI session that runs in a native desktop app instead of a browser tab.

If it's both, say so out loud in your evaluation, because it changes the shortlist. Plenty of companies genuinely have both problems, and pretending one product covers both wastes a quarter. What we will argue is this: on the human side, you can buy a point product and still need a gateway, still need endpoint DLP, still need tenant login control. Or you can get the gateway, the endpoint, the SaaS tenant, and the AI prompt in one console with one agent.

Deployment speed backs that up. One Fortune 100 rollout moved at thousands of devices a week through a silent Intune deploy, and Outreach Health had its 34 offices covered inside a week. Their words, not ours: "It's not just great, it's dope."

Book a 20-minute demo

Bring your AI tool list and we'll show you what Shadow IT discovery, Cloud Application Control, and Dopamine DLP catch on a real device. Or start a free production trial and see your own AI Usage Analytics inside a week.