Security and Compliance

Do not trust the agent. Control the action path.

IntentFrame assumes the agent may be wrong, tricked, or compromised. The rules live outside the agent, sensitive actions run through a governed path the agent does not control, and every proposed action is checked before execution.

The agent is untrusted input.

IntentFrame does not rely on the agent to follow instructions. The agent's request is treated as evidence for evaluation, not authority. It can propose an action, but it cannot approve the action or directly touch the real system.

  • Rules live outside the agent.
  • Sensitive actions run through a governed path the agent does not control.
  • Every action is checked before it runs.
  • Unclear means blocked or sent to human review.
  • Every decision can leave evidence.

The Core Distinction

Structural prevention, not surveillance.

Monitoring watches agents as they act. IntentFrame prevents them from acting without validation. The agent already has capability when surveillance notices something wrong.

Structural preventionSurveillance (monitoring)
Credential accessOnly the validated execution path has credentialsAgents have credentials
Attack timingPrevents execution capability from being misusedDetects attacks after capability exists
Defense typeArchitectural — novel attacks still hit the boundaryPattern-based — can miss novel attacks
Response modelProactive enforcement before executionReactive alerting
Security outcomeStops what should not happenLogs what happened

Monitoring tells you what happened.

IntentFrame controls what is allowed to happen.

Zero Hopeful Trust

Hope is not a security model.

Most agent deployments still rely on hopeful trust: hope the developer scoped permissions correctly, hope the agent interprets instructions safely, hope prompt injection does not slip through, hope side effects were anticipated.

  • Hope the developer scoped permissions correctly
  • Hope the agent interprets instructions safely
  • Hope prompt injection does not slip through
  • Hope side effects were anticipated

Hope is not a security model.

IntentFrame replaces hope with an enforced boundary.

What IntentFrame is built to stop.

Prompt injection consequences

We do not claim to stop the model from being tricked. We stop the unauthorized action that results from the trick.

Unauthorized money and data actions

Refunds, payments, account changes, or record updates that break your business rules are blocked before they run.

Scope violations

Agents acting outside approved accounts, recipients, targets, or action types are blocked.

Semantic deception

An action whose stated reason does not match what it would actually do is caught by cross-checking the request against its own details.

Destructive actions

Deleting or altering records and data — and, where an agent has system access, files and infrastructure — can be governed before execution.

Audit gaps

Every decision can produce a record explaining what happened and why.

We do not claim to stop prompt injection. We stop the resulting action.

Prompt injection can still trick the model. IntentFrame's job is to prevent the agent from turning that compromised reasoning into an unauthorized action.

  • If an attacker convinces the agent to ignore policy, that instruction is still just text.
  • If the agent proposes an unsafe action, IntentFrame evaluates the action against trusted rules.
  • If the action violates policy, it does not run.

We cap damage at the boundary. We do not pretend to eliminate the cause.

Fail-Closed

Fail-closed by default

Unclear means blocked or sent to human review — never silent approval. The safe default is to stop execution when validation cannot confidently allow an action.

Validation unavailable? Execution halts.
Intent unclear? Execution halts.
Pattern unexpected? Execution halts.
System under uncertainty? Execution halts.

Security is the default state.

Every decision needs a receipt.

For production agents, logs are not enough. Security and compliance teams need to reconstruct what the agent asked for, which rule applied, why a decision was made, and whether the action executed.

  • Actor / agent identity
  • Proposed action
  • Action parameters
  • Business context
  • Policy version
  • Decision: ALLOW or BLOCK
  • Reason / rationale
  • Timestamp
  • Executor result
  • Hash / chain metadata where configured

The audit trail should answer: what did the agent decide, what did IntentFrame allow, and on whose authority?

For Security Reviewers

For security reviewers: how the boundary actually works.

The summary above is the honest version. This section adds the detail a security team needs to evaluate the model, including what runs without AI, where AI is involved, and what is explicitly out of scope.

The claim we want you to test+

We make one falsifiable claim: an action that violates your policy does not execute through the governed path, even if the agent is fully compromised. That claim is meant to be attacked. We publish adversarial test results rather than asking you to take it on faith.

What is enforced without any AI+

Hard limits are checked deterministically before any model runs: amount caps, allowed recipients and accounts, allowed action types, required fields, and obviously dangerous operations. Execution side deterministic checks may also run immediately before an allowed action is carried out. If these checks fail, the action is blocked regardless of what any AI thinks.

Trusted vs. untrusted components+

Trusted: your written policy and the deterministic checks. Untrusted: the agent and anything it ingests, including its own reasoning and tool output. Partially trusted: the AI review steps, which can interpret an action but cannot grant authority — their output is still checked against deterministic policy before anything runs.

How injected instructions are contained+

IntentFrame does not try to detect every injection. Instead, malicious text the agent absorbs can only ever become a proposed action, and every proposed action is checked against trusted policy that the agent cannot edit. The reason given for an action is also cross-checked against the action's real details, so 'process a routine refund' that actually wires money to a new account is caught.

What is in scope and what is not+

In scope: blocking individual unauthorized actions, enforcing business limits, separating reasoning from authority, and producing per-action evidence. Out of scope: making the base model correct, stopping injection from changing the model's internal state, detecting abuse spread across many individually-allowed steps, and any action that bypasses the supported IntentFrame path entirely.

Shipped hardening and known gaps+

Shipped: deterministic limits, separated AI review, fail-closed defaults, and tamper-evident audit records. Known gaps we are honest about: policy quality is still your responsibility, multi-step 'slow' abuse is an active area of work, and AI review adds latency and inference cost on the actions that require it. We would rather tell you the edges than oversell them.

Runtime evidence for agent actions.

IntentFrame is not a compliance certificate. It is an enforcement and evidence layer that helps teams satisfy review requirements around policy, access, decisioning, and auditability — increasingly a hard requirement as agent regulation takes effect.

EU AI Act

Per-action records, control evidence, and runtime governance for agent decisions as enforcement timelines arrive.

HIPAA and healthcare

Enforcement and access controls that hold at the action layer, not just in a system prompt an agent can ignore.

SOC 2

Control evidence around authorization, change management, logging, policy, and review.

ISO 42001 / NIST AI RMF

Evidence for AI governance, risk management, monitoring, and accountability.

Financial services

Runtime control, decision evidence, and separation between recommendation and execution.

Compliance requirements vary by organization, jurisdiction, and deployment. IntentFrame provides controls and evidence that your team can evaluate as part of your compliance program.

Security posture depends on where it runs.

IntentFrame supports multiple deployment profiles so teams can choose the right data boundary.

Cloud

Dedicated managed deployment. Fastest path for teams that can use a managed endpoint.

Private Deployment

Runs inside your cloud or VPC. Data and credentials stay in your environment.

Self-Hosted

Customer-operated runtime for maximum infrastructure control.

On-Prem / Air-Gapped

Scoped during architecture review for regulated or disconnected environments.

Where your data goes during a decision.

Hard-limit checks run locally/cloud and send nothing out. When an action needs AI review, the action details and relevant context are sent to the model provider configured for your deployment. In private, self-hosted, and air-gapped deployments, that review stays inside your environment, so sensitive data never has to leave your boundary to be governed.

What we protect and what we do not.

Protects

  • Individual actions, checked before they execute.
  • Business policy boundaries like amounts, recipients, and accounts.
  • Direct agent access to credentials and sensitive systems.
  • Unauthorized actions that require judgment about meaning.
  • Auditability of every allow and block decision.
  • Per-action compliance evidence.

Does not protect

  • It does not make the underlying model correct.
  • It does not prevent prompt injection from changing the model's internal state.
  • It does not, on its own, catch abuse spread thinly across many individually-allowed actions.
  • It does not govern actions taken outside the supported IntentFrame path.
  • It does not replace identity, network, endpoint, or application security.
  • It does not certify your organization as compliant by itself.

Bring IntentFrame into your security review.

For enterprise evaluations, we can support architecture review, deployment scoping, policy design, audit evidence mapping, and security documentation.

  • Agent stack and tool calls
  • Credential and executor model
  • Policy authoring and versioning
  • Audit storage and retention
  • Deployment profile
  • Data residency requirements
  • Compliance mapping
  • Support and operational process