Opsin raises $21 million on a permissions problem that agents make urgent

By
CTOL Staff Reporter
1 min read

Opsin tripled its funding from a $7 million seed in April 2025. Its only published statistic measures tool access and approval gates, settings that a platform vendor can change.

Opsin, a security start-up, has raised $21 million to sell protection against AI agents that misbehave inside large employers. SYN Ventures and Race Capital led the round, according to Opsin's release. Race Capital also led the seed, so the product pitch has changed more than its backer list.

Opsin sells control over who and what can reach each set of data. That is an old access-control problem. Agents make it more urgent, but the control remains the same.

An agent inherits the permissions of the person who built it, then acts on them without a human in the loop. A platform vendor that sets those permissions by default can capture most of the product's value.

The company's own words, eighteen months apart, show the continuity. In its April 2025 funding post, announcing a $7 million seed, Opsin wrote that "GenAI hasn't created new risks. It has simply made existing ones impossible to ignore." The problem it named was "the lack of visibility into who can see what, and the blast radius of unintentional access". Its 2026 platform is "built to detect and contain AI agent misbehavior". The new round is three times the seed.

The 2026 release supplies one number, and the argument turns on it. Opsin analysed agent activity across its enterprise customers. It says 93% of "misaligned tools" connected to agents had "full, unrestricted access" to the systems they were connected to. "Just 1 in 200 required human approval before taking action."

Both figures describe permissions. The first measures how much access a tool was granted. The second measures whether an approval gate was switched on.

One in 200 is 0.5%, so about 99.5% of those misaligned tools could act without anyone signing off. The sample is Opsin's own customers. They bought a security product because they were already worried, so the figure describes that group, not the market. Opsin does not define what makes a tool misaligned, so the count is hard to test.

Opsin's answer is that permissions are only the start. James Pham, its co-founder and chief executive, says in the release that "knowing what an agent has permission to access isn't enough. Security needs to understand what the agent is trying to do." The release describes a system that catches "when behaviors deviate from their intended purpose". Pham's founder post goes further. It says Opsin watches run-time behaviour, such as an agent accessing data outside its role, and can quarantine an agent or block an action.

Agents chain actions and run continuously, so a permission that was harmless to a person can become a hazard at machine speed. That explains why the same control now carries a larger budget. Run-time blocking is a real feature, and it is the part a vendor cannot copy in a day. The evidence Opsin chose to publish, though, is a count of grants and gates. A setting fixes both when an agent is created.

The company's own boilerplate lists the job as helping organisations "discover AI agents, identify oversharing and excessive access, prioritize risk, and enforce security controls". Two of those four verbs, identifying excessive access and enforcing controls, are permissions work. Race Capital's Alfred Chuang says Opsin "connects what an agent is supposed to do with what it actually does". The first half of that sentence is a configuration.

Intent monitoring uses the same inputs. Opsin says its platform connects an agent's purpose, permissions, data access, tools and activity. A purpose is only as visible as the vendor's own console makes it. If a platform vendor records an agent's owner, scope and approval status natively, the independent layer keeps the cross-vendor view and little else.

Microsoft, OpenAI, Anthropic and Google ship the tools in which employees build agents. Opsin's platform covers agents built "across Microsoft Copilot, Claude, ChatGPT Enterprise, Gemini", and other approved tools. Its neutrality across vendors is the selling point. Its dependence on them is the weakness. Opsin can see across vendors, but it cannot change what any one vendor chooses to ship.

The customer list fits that story. Opsin says it has nearly 20 enterprise customers, including Culligan, Wellstar Health System and UiPath, across healthcare, financial services and manufacturing. Culligan was already named in the 2025 post, where Opsin credited its product with an 80% cut in sensitive data exposure. The same buyers who paid to stop Copilot oversharing are now being sold agent governance. They know the product, which makes them the cheapest place to sell the next one.

For a head of security or identity at a company rolling out Copilot Studio or ChatGPT Enterprise agent builders, the first route is a default rule. No agent is published with write access to financial or HR systems without a named owner and an approval step. It slows the rollout for employees who build agents. It also applies only to agents published after the setting changes.

The second route is a governance platform that finds and scores agents once they exist. It costs a licence and a second inventory that duplicates the identity system. It is the only route that reaches agents already deployed. How many agents are live, and how much write access they hold, sets the order.

On Opsin's own statistic, the gap sits in the approval setting. Turning it on is cheaper than buying a vendor. By 31 March 2027, Microsoft will have added a Copilot Studio admin console setting that blocks any agent with write access to financial or HR systems until a named owner and approver are recorded. That will decide whether $21 million bought a product or a feature.

You May Also Like

This article is submitted by our user under the News Submission Rules and Guidelines. The cover photo is computer generated art for illustrative purposes only; not indicative of factual content. If you believe this article infringes upon copyright rights, please do not hesitate to report it by sending an email to us. Your vigilance and cooperation are invaluable in helping us maintain a respectful and legally compliant community.

Subscribe to our Newsletter

Get the latest in enterprise business and tech with exclusive peeks at our new offerings

We use cookies on our website to enable certain functions, to provide more relevant information to you and to optimize your experience on our website. Further information can be found in our Privacy Policy and our Terms of Service . Mandatory information can be found in the legal notice