
The rules did not pause for AI agents
Banks, insurers, health plans and government agencies want the same AI agents everyone else is adopting: coding assistants, document summarizers, research agents. An AI agent is software that works for an employee: it reads, decides and acts using an AI model. The rules these organizations live under were written before agents existed, and they still apply.
- Customer, patient and citizen data has to stay where the rules say it lives.
- Each system gets only the access it needs, and that now includes software acting for a person.
- Every action has to be recorded, and the record has to hold up when an examiner asks.
- The most sensitive networks have no internet at all. In an air gap, data moves "only manually, under human control."¹
The gap is real: IBM found that 92% of organizations with an AI-related breach lacked proper AI access controls.²
In a regulated company the question is never only "is the AI useful?" It is "can we show who did what, with which data, under which rule?"
What Argent does: one checkpoint, inside your walls
Argent Strategic Intelligence sits between your AI agents and the models and tools they use, and decides each request before it goes out. It runs inside your own environment, detection included, and it can run with no internet at all.
| What the rules ask | How Argent handles it |
|---|---|
| Keep sensitive data in | Reads each request first and replaces Social Security, card and bank account numbers with a typed marker such as <US_SSN_REDACTED> before the model sees them, or blocks the request, per your policy. See sensitive data redaction. |
| Use only approved AI | The approved-model list starts empty: a model nobody approved is refused. Models can run on your own servers. See model control. |
| A person answers for every agent | Every agent is tied to a named human owner. When the owner is deactivated, the agent is refused on its very next request. See identity tethering and owner offboarding. |
| Least access | Tools are approved one by one, and a new or changed tool waits for an admin's decision. The agent never holds the provider's key. See agent reach control and keyless agents. |
| Prove it later | Every decision goes on a signed record, each row linked to the one before. See one-click audit evidence and session replay. |
| No internet required | Installed from one signed bundle carried in on media. See air-gapped install. |
| Do not break the work | Starts in watch mode: Argent records what it would have done while requests pass unchanged. See watch before you enforce. |
A walk-through: a health plan's claims agent
An example: a health plan runs its claims systems on a sealed network. A claims analyst wants an AI agent to summarize claim files with a model running on the plan's own servers.
- The release is carried in. One signed bundle arrives on removable media. The import checks the release signature, the checksum list, every file and each image's digest, prints
FATAL:and stops on any mismatch, then loads the plan's own registry. One setting,global.imageRegistry, points the install at it. - Watch first. The claims team starts in watch mode. Argent records what it would have done and forwards each request unchanged, so the security team sees real traffic before anything is enforced.
- The agent gets an owner. The agent is registered to the analyst by name. It calls the gateway with its own agent key and never holds the model provider's key.
- A request carries a member's Social Security number. Once enforcement is on, Argent replaces the number with
<US_SSN_REDACTED>before the model sees it. The record names the kind and the count, never the value. - The model answers, inside. The cleaned request goes to an approved model on the plan's own servers, such as Ollama or vLLM. A model nobody approved is refused. Each decision goes on a signed record.
- The agent asks for a new tool. The new tool waits for an admin's decision before the agent can use it.
- The analyst leaves. The analyst's account is deactivated, and the agent is refused on its very next request.
- The examiner arrives. A read-only auditor builds an evidence package in one click. The examiner checks it on their own machine and sees
TENANT PACKAGE OK. The auditor can also replay any agent session prompt by prompt.
What the rules already ask for
Each rule below applies to a named audience. Argent gives the security team the controls and the signed records these rules ask about, for every agent that works through it.
| Rule | Who it applies to | What it asks | Where Argent helps |
|---|---|---|---|
| HIPAA Security Rule, 45 CFR 164.312(a)(1) | Health plans, health care clearinghouses and providers that transmit health information electronically, and their business associates³⁴ | Access "only to those persons or software programs that have been granted access rights"³ | Each agent has its own identity and owner; tools and models are approved one by one |
| HIPAA Security Rule, 45 CFR 164.312(b) | The same health organizations³⁴ | Mechanisms "that record and examine activity" in systems that contain or use electronic protected health information³ | A signed record of every request and decision, with session replay |
| NIST SP 800-53 Rev. 5, SC-7(5) | US federal information systems at the moderate and high baselines⁵ | Deny network traffic by default and allow it by exception⁵ | Argent's own services open no outside connection by default; anything that leaves is something your team routes or switches on |
| NYDFS AI guidance (October 2024) | Financial services companies regulated by the New York Department of Financial Services⁶ | Adds no new requirements; points to limiting access to what the job needs, inventorying every system that uses AI, and monitoring for unusual queries⁶ | A live list of every agent with its owner, and an access review of every grant that downloads as CSV |
For the banking view, see Argent for banking; for agencies, Argent for the public sector; for health, Argent for healthcare.
Capabilities and integrations involved
Argent plugs into the systems a regulated company already runs. Each one below has its own page.
- AI models on your own servers: Ollama, vLLM, LM Studio, NVIDIA NIM and Hugging Face TGI.
- Outside AI providers, where you allow them: Amazon Bedrock, Azure OpenAI and the rest of the integrations.
- Identity and owners: Microsoft Entra, Okta, Ping Identity, SailPoint, SCIM and LDAP and Active Directory.
- Secrets, kept in your vault: HashiCorp Vault and OpenBao and CyberArk Conjur.
- Sensitive-data detection: Presidio, shipped inside the signed release with its language model built in.
- Agents: Claude Code, Codex CLI, GitHub Copilot CLI and frameworks such as LangGraph.
- Install: the air-gapped bundle, Kubernetes with Helm or Docker Compose on one server.
Where it runs, and how a regulated team starts
The same signed release installs on your own servers with no internet, in a private cloud, or in your own account in a public cloud. Compare air-gapped on premises, private cloud and public cloud VPC on the deployment options page.
- Carry in and check. Bring the signed bundle across, run the import, and load your own registry.
- Watch. Point agents at the gateway and routes at your models. Argent records what it would have done and forwards each request unchanged.
- Enforce, area by area. Data checks, models, tools and owners switch on independently, when you decide.
- Goals: sovereign AI, compliance evidence, audit and accountability, stop data loss, agent identity and ownership and all goals.
- Other use cases: AI coding agents, customer-facing AI assistants, enterprise copilots and MCP and agent-to-agent tool access.
- Industries: banking, insurance, healthcare, public sector, military and defense and all industries.
Take the walkthrough, browse all capabilities, or get the data sheets.