Skip to main content
AI Agent Security · Evidence-Bound Assurance

Your AI has permission. That doesn’t mean it’s safe.

HAIEC shows what your AI can really reach, what it actually did, and where the evidence stops — so you can prove what’s true and be clear about what isn’t.

Real-World Patterns

This is not hypothetical.

Publicly disclosed incidents keep repeating the same shape: individually legitimate capabilities connecting into an unintended consequential path.

Public input → private data → public output

A malicious public issue influenced an AI agent that could read private repositories and create public pull requests. Individually valid capabilities formed a data-leak path.

Trusted tools can create an untrusted sequence.

Illustrative pattern. HAIEC evaluates whether a comparable path exists in your system.

Publicly disclosed pattern · Invariant Labs, May 2025

Email → internal context → external leak

A crafted email could influence an enterprise AI assistant and cause internal data to leave through an external channel, without the user intentionally sharing it.

The input itself can become an instruction source.

Illustrative pattern. HAIEC evaluates whether a comparable path exists in your system.

Publicly disclosed pattern · Microsoft Security Insider, 2025

Agent → production data → destructive action

An AI coding agent deleted app database data during development; the platform later strengthened development/production isolation.

Natural-language intent is not the same as enforced authority.

Illustrative pattern. HAIEC evaluates whether a comparable path exists in your system.

Publicly disclosed pattern · Replit, July 2025

These are publicly disclosed patterns, not HAIEC findings. HAIEC does not claim to have detected or prevented any named incident. The assurance question is whether a comparable influence, reach, authority, and consequence path exists in your system, and what the evidence establishes about it.

System Constellation

See the paths a checklist cannot show.

HAIEC expands the agent into the system around it: what can influence it, what it can reach, what controls the path, whether another route leads to the same effect, and what the evidence actually proves.

Influence

User input, knowledge, memory, tool output, peer agents, runtime configuration

Reach

Tools, APIs, data, identities, cloud, workers, external services

Control

Authentication, authorization, tenant scope, approval, validation

Paths

Canonical path, async workers, direct tool calls, webhooks, retries, legacy fallbacks

Consequence

Data change, money movement, access change, production change, external communication

Evidence

Established, partial, unknown, or not assessed — with inspectable proof and an explicit frontier

See the evaluated Kestrel System Constellation
Sample environment · illustrative data
Static Sample
Northstar Commerce Operations Copilot
AI-assisted refund processing
Request
User / application input
Policy
System instruction
AI Agent
Model call
Action
stripe.refunds.create
Policy
Authentication
Action
Canonical path
Consequence
Stripe refund
Potential consequence
Permission ≠ Delegation
Permission can be valid while delegation of the consequential choice remains unproven.
Evidence & Authority

What does the evidence actually prove?

No vulnerability does not mean no consequential capability. A perfectly valid tool, called through a valid API, using a valid service identity, can still produce an action the organization never intended the AI to choose.

Permission is not delegation. Security controls answer whether an identity is permitted to do something. They do not necessarily establish whether the AI was delegated the consequential choice it made.

Permission

An identity policy that allows an action. Establishes what is permitted, not what was authorized by a human decision.

Delegation

Whether a human or process with authority actually delegated the consequential choice to the AI. Distinct from permission. Not always established by available evidence.

HAIEC does not claim delegation violations are currently proven unless evidence establishes them. The distinction between permission and delegation is a structural insight, not an automatic finding.

1. Requested

What the request asked the AI to do

2. Policy Authorized

What policy and configuration permit

3. Effectively Granted

What credential evidence establishes

4. Code Capable

What the application can technically do

5. Observed

What runtime evidence establishes was observed

These are evidence questions, not a causal proof chain. A gap between any two is a finding, not a safe state. Not every question currently has a native evidence producer.

Authority — what this agent can do, by evidence planeIllustrative trace

Asked for

reroute_traffic on netops.example.com
query_records on customer-db.example.com
export_report to s3://ops-reports

Allowed by policy

narrower than the plane above
reroute_traffic on netops.example.com
query_records on customer-db.example.com

Granted by credentials

narrower than the plane above
reroute_traffic on netops.example.com

Possible in code

no declarations in evaluated source

Seen in evidence

includes a host nothing authorized
reroute_traffic on netops.example.com
reroute_traffic on attacker-node.example.comdivergence
query_records on customer-db.example.com

The planes are independent evidence planes — what was asked, what policy allows, what credentials grant, what code can do, and what was observed. They are not a guaranteed sequence.

What was asked, what policy allows, what credentials grant, what code can do, and what was observed — side by side. See the evaluated Kestrel report →

Permission is not delegation. Delegation is not authority. Authority is not observed behavior. HAIEC evaluates each plane and shows where they diverge.

Capability chain

What a task requires vs. what the agent can actually do. Unmet requirements, unobserved grants, and capabilities exceeding the request render as distinct gap types.

Delegation chain

Whether authority actually narrowed at each hop — scope monotonicity and chain continuity. A break is surfaced, never averaged away.

DAI comparison

Four independently-evidenced facts per surface — delegated, code-capable, granted, observed — compared side by side. Divergence is shown, not reconciled silently.

Alternate paths

Expected path

Request → approval → action. The mediated route reviewers expect the AI to take.

Alternate path

A background worker, fallback handler, or direct tool path that reaches the same effect through a different route.

Finding one approval does not prove every route to the action is protected.

How HAIEC Works

From repository to decision.

A bounded lifecycle for evidence-bound assurance of consequential AI systems.

Define

Select the AI system and assurance question

Connect

Connect evidence sources: repo, APIs, and infrastructure

Collect Evidence

Source, identity, access, policy, capability

Assure

Deterministic evaluation within explicit scope

Verify

Decision Receipt with SHA-256 integrity

Monitor

Ongoing evidence from supported runtime sources

Monitor is bounded to supported evidence and runtime sources. Not every connected asset has a native evidence producer. Evaluated Scope is currently structured at the contract level; structured scope binding is under development.

Capability state as of build v1.0.0 — auditable in the repository.

Trust Signals

Standards-informed. Evidence-decided.

Grand Challenge Winner

MunichTech EXPO 2026

Evaluated against KestrelVoice, a real multi-tenant AI SaaS: 44/44 action paths met criteria, 1,322 evidence traces, 555 verified checkpoints — jury score 8.7/10.

Deterministic evaluation

No LLM in the canonical detection or assurance path — same input, byte-identical digests.

Open evidence formats

STIX 2.1 indicators for SOC interchange; OTLP intake for runtime telemetry.

Integrity by construction

SHA-256 digests on receipts and case artifacts — recomputable, never edited.

Public samples

Evaluated sample systems are public — read the report, the constellation, and the forensic case yourself.

A Different Question

Different questions. HAIEC connects the answers.

Buyer questionScannerGovernanceIAMHAIEC
Does the path exist?PatternPolicyAccessPath + consequence + evidence
What influences the choice?—Partially—Yes — influence chain
What's the consequence?—Policy text—Yes — consequence graph
What does the evidence prove?Findings countCompliance %Access logsEstablished / partial / unknown + frontier
Where does proof stop?———Yes — evidence frontier

These tools answer different questions — this is not a claim that any of them are wrong. HAIEC connects their answers to the consequential path.

Evidence & Framework Support

Mapped to standards. Not defined by them.

HAIEC maps evidence to recognized frameworks and regulations. This is evidence mapping, technical evaluation, readiness, and assurance support. Framework mapping is not the same as assurance. A green or source-established fact is not the same as safe.

AR-01

Instruction-Like Pattern in Model Context

ATLAS AML.T0051ATLAS AML.T0051.001ASI ASI01NIST MEASURENIST MANAGE

AR-22

Prompt / Context Attack Signal

ATLAS AML.T0051ATLAS AML.T0051.000ATLAS AML.T0054ASI ASI01NIST MEASURENIST MANAGE

AR-69

Handover Outside Declared Scope

NIST MANAGE-2.43GPP TS 38.3003GPP TS 23.502

Framework mapping ≠ certification. Standards inform. HAIEC evaluates.

Browse all framework mappings

HAIEC supplies evidence and analysis. The qualified firm, CPA, certification body, or assessor retains professional judgment and issues the applicable attestation, certification, or assessment outcome. Framework mapping does not equal assurance.

Evidence-Bound Assurance for Consequential AI

What can this AI actually cause?

One bounded system. One primary assurance question. Explicit evidence boundary. Decision-grade technical and executive output.