An agent can do more than produce an answer
An AI agent can read a document, call a tool and act on a business system. A security problem can travel through those steps before it becomes visible in the final response.
Armorer Labs follows that path through the agent. It maps the tools, permissions, identities and data the agent can reach, then tests whether an attack can cause a real failure. [1]
The company builds continuous security verification for teams deploying agents in regulated industries. Finding the issue is the start; Armorer also prepares a fix and checks that it works. [1][2]
Find the path an attacker can use
Armorer examines the agent’s actual environment: what it can read, which actions it can take and where its permissions stop. [1]
Its testing combines attack simulation with verification against the target. The goal is a reproducible failure that an engineer can investigate, tied to the affected tool or permission. [1][3]
That gives the team a concrete next step. Armorer can propose a small code change or a runtime policy, then test the revised system again. Engineers retain approval over the changes. [1]
Verify the fix and the normal job
Consider a finance agent that reads an external document and then uses business tools. A malicious instruction inside the document could try to redirect that access.
Armorer can test whether the agent crosses a boundary, reproduce the failure and prepare a narrower permission or policy. It then checks whether the attack is blocked and the intended business task still works. [1][3]
That second check matters to the product’s design. A fix needs to remove the risk while keeping the agent useful to the people who rely on it. [1]
Armorer keeps the finding, proposed change and verification together so an engineer can review the chain before approving a deployment. [1][3]
Carry the policy into production
Armorer Guard adds checks at runtime. Actions can be allowed, held or denied against the team’s policies while the agent is working. [1]
The service supports cloud, on-premises and air-gapped deployments. Teams can choose an environment that fits where the agent and its data already run. [1]
Findings and controls can be mapped to security frameworks for review and audit work. Those mappings help the team explain its controls alongside the technical evidence. [1][3]
Gaetan Licci, Director of Engineering at Just Eat Takeaway, highlighted the value of resolving a security problem rather than adding another finding for the team to triage. [1]
Built by people who secured agents at AWS
Daniel Begimher and Cristian Leo worked together at AWS on security agents and their evaluation before founding Armorer Labs. [2]
Their work included open security tools and benchmarks, including ASH, ThreatForest and SIR-Bench. Armorer brings that experience to the agents other companies are building. [2]
The company joined Y Combinator’s Fall 2026 batch. Teams can book a demonstration around their agent, deployment environment and security requirements. [1][2]
Armorer’s direction is continuous verification: test the agent, fix what fails and keep enforcing the boundary as the system changes. [1]
Where Armorer Labs ranks
Armorer Labs isn't in a numbered ranking yet. Products get a position once enough companies compete on the same task.
Current rankings. Positions come from public evidence, and no company can pay for a place. How we rank
The bigger picture
A company deploying an agent needs to know which actions are safe and what happens when they are not. [1]
Armorer Labs is building the testing, verified fixes and runtime controls that make those boundaries part of the agent’s operation. [1][3]
Written by SOTA2 from Armorer Labs’s public materials. Customer experiences and product plans are reported by Armorer Labs.