Agents over humans
Agents burn nearly 5× the tokens people do on OpenRouter.
The security
platform for
AI agent work
Enterprises are shipping agents. Security still has no system of record for what those agents can do. Start with one workflow: enforce signed policy before privileged actions run, and keep a verifiable record of what was allowed, reviewed or blocked.
Why now
Agents already do most of the AI work. Governance does not.
Observed agent traffic is burning far more tokens than humans, while most organizations still lack durable policy and evidence for what agents can do.
Since February
Agentic token usage is up about 14× since February 2026 on OpenRouter.
Adoption outpaces governance
In IBM's survey, 77% of organizations said AI adoption is moving faster than their governance capabilities.
Oktsec in production
Policy before action in a regulated clinical workflow.
Certuma runs AI reasoning inside deterministic safety gates, with physician sign-off on every plan. Oktsec governs the agent path: explicit policy before privileged work, auditable evidence after.
certuma.ai · Doctor-verified clinical careClinical workflowAutonomous clinical infrastructure
AI reasoning inside deterministic safety gates, with physician sign-off on every plan.
Oktsec in productionGoverned agent work in regulated healthcare
Policy before action. Auditable evidence after.
Policy before action.
Evidence after.
Expand agent automation without widening unchecked authority. Control applies explicit signed policy before privileged work runs, routes exceptions for review and returns evidence security teams can verify.
Enforcement is deterministic. A model can recommend. It cannot authorize.Policy assignedThe company defines which actions, tools, paths and destinations are allowed.
Signed instructionsA signed policy bundle is distributed to enrolled environments.
Local enforcementOktsec applies the decision before privileged work runs.
Evidence returnedThe request, decision and policy version are reported back.
Exception reviewOut-of-scope actions are blocked or held for review.
Approved environments
Environments
PolicyAssigned
ExecutionRunning
EvidenceVerified
ExceptionsNeeds review
prod-apiServerRunningProduction rules
agent-runnerWorkstationRunningAgent runtime
ci-pipelineCINeeds reviewCI rules
policy-runtimeContainerRunningMCP rules
Recent evidence
prod-apiRead approved repositoryAllowed
agent-runnerCall approved MCP toolAllowed
ci-pipelineExternal network actionNeeds review
Mapped to the frameworks security teams already use.
Connect Oktsec controls, assessments and evidence to established security and AI risk frameworks.
Start with one workflow. Expand without losing control.
Begin with one consequential path. Test it against its real execution surface, govern it at runtime, then operate the same boundary across teams and environments.
- 01 / Assessment
Validate the workflow
Exercise the real execution path. Get reproducible findings, evidence and a remediation plan.
Scope an Assessment - 02 / Control
Govern actions at runtime
Apply explicit policy before privileged work runs, then record the decision and its evidence.
Review a control boundary - 03 / Cloud
Operate across environments
Manage approved environments, policy versions, exceptions and evidence from one operating view.
Explore Cloud
Security in the path agents already use.
Oktsec sits between agent clients and the systems they can change. Teams keep their runtime, MCP servers and infrastructure while adding one enforceable boundary for tool calls, policy and evidence.
Claude Code · Claude Desktop · Cursor · VS Code · Codex · ChatGPT · Gemini CLI · Grok · agent CLIs
Teams keep their agent tools; Oktsec governs the action path through wrappers, hooks, MCP gateway and authenticated HTTP surfaces.
MCP servers · repositories · internal APIs · cloud accounts
Actions stay in customer-controlled environments while policy is applied before privileged work runs.
Policy · evidence · exceptions · reporting
The cloud control plane manages governance, verifies evidence and gives teams one operating view.
The control gap is already operational.
Enterprise adoption is moving faster than governance. Once agents use tools, credentials and the live internet, model behavior alone is not an authorization boundary.
Security requirement Policy must be enforced where the action happens.
See how Oktsec enforces policy01Enterprise scale
Governance is trailing agent deployment
A June 2026 study of 2,000 technology executives found that AI adoption is moving faster than governance. The operational question is which actions deployed agents are authorized to take.
02Operational scale
Security review is moving to agent scale
OpenAI reports that Codex Security scanned more than 30,000 codebases and that human reviewers marked over 70,000 findings as fixed. Scope, validation evidence and review now have to scale with agent capability.
03Model behavior
Model behavior is not an authorization boundary
Controlled simulations across 16 leading models found harmful insider behavior in at least one model from every developer tested. Runtime authority still needs an enforceable boundary.
Built from the security failures agents expose in practice.
Oktsec turns original security research into deployable controls for agent identity, tool use, execution boundaries, third-party dependencies and verifiable evidence. Findings from agent tooling, developer systems and cloud infrastructure inform what the platform tests and controls next.
ECOSYSTEMAAIF Ambassador · Linux Foundation
OPEN SOURCEAguara and nanostack, on GitHub
RESEARCH ACCESSOpenAI Daybreak Blue · Anthropic Cyber Verification
RESEARCH250+ findings reported to Google, Microsoft, Stripe and AWS
Founded by Gustavo Aragón, 20+ years across security, product and regulated fintech.
Meet the founderResearch that becomes product.
Oktsec's controls are shaped by security findings reported through major programs and direct disclosure channels.
Google
Microsoft
Stripe
Cloudflare
AWS
Mercury250+security findings reported through VRP, MSRC, HackerOne and direct disclosure
Expose the attack paths in your AI agent workflow.
Find exploitable agent paths before production. Oktsec's purpose-built assessment harness exercises real execution paths across agents, MCP servers, tools, repositories, credentials and network access, producing reproducible findings, evidence and remediation.
01reviewed risk recordEvidence, severity, remediation and an executive summary in one deliverable.
Findings engineering can reproduce and leadership can act on.
Each Assessment produces reviewed findings, executable evidence and a practical remediation plan for the workflow your team scoped.
See how Assessment worksAssessment report
Finding · OKT-142
HighPrivileged tool call outside approved boundary
A coding agent attempted to pass a secret-bearing file to an unapproved external destination. The path was reproduced and the boundary decision verified.
OverviewEvidenceImpactRemediation
agent → coding-assistant
tool → shell.exec
request → exfiltrate .env via curl
boundary → deny: egress + secrets
decision → block + review
evidence → integrity verifiedField notes from the audit trail.
Primary research and reproducible labs on authorization, MCP security, agent identity, supply chain and public incidents where agent work moved beyond the intended boundary.
Explore Oktsec ResearchWhat security teams need to know before deployment.
Where does it run?
At the enforcement points between agents and the tools, code, data or network destinations they can change.
Who makes the decision?
For actions routed through Oktsec, explicit deterministic policy decides whether privileged work is allowed.
What evidence returns?
A signed, tamper-evident record of identity, request, decision, policy version and integrity state.
How do we start?
Choose one workflow that touches code, tools, credentials or infrastructure, then place Control around that boundary.
How does Oktsec handle identity and delegation?
Every policy decision is tied to a principal. Oktsec supports signed agent messages, bearer-authenticated gateway clients, reported actors for sub-agents, scoped constraints and cryptographically verifiable delegation chains.
Does Oktsec need inbound access to our environments?
No. Execution stays inside the customer environment. Approved environments can pull signed policy, apply it locally and return evidence to the managed control plane. Private, isolated and offline control-plane modes remain planned until they are shipped and verified.
Start with one consequential agent workflow.
Choose the path that touches code, tools, credentials, customer data, network access or infrastructure. Define its boundary, test it and preserve evidence before expanding scope.