Running release
ReadingThe build serving production right now, signed.
Trust Center
ALEETH exists to put autonomous AI under institutional control. It would be a contradiction to run our own platform any other way. The live record below is read from production as the page loads.
AT REST
AES-256
database, including secrets
PROVENANCE
Ed25519
append-only, anchored
ISOLATION
Per-tenant
customer tokens, API-authorized
AUDIT
Append-only
database-trigger enforced
The Live Record
The build ICA is running, the transparency log ICA seals receipts into, and authenticated-route coverage for ALEETH's own platform routes, as ICA publishes it. Your browser reads each one from ICA directly. The build and transparency-log figures link to their raw records.
Running release
ReadingThe build serving production right now, signed.
Transparency log
ReadingReceipts are sealed into Merkle trees. On a schedule, ICA time-stamps its latest anchored root with an RFC 3161 service. The inclusion proof for a receipt you hold can be checked offline.
Authenticated-route coverage
ReadingThe share of ALEETH's own platform API routes with side effects that require authentication or another recognized control (tool gateway, shared secret or credential), from a static scan of ICA's code checked in CI, as ICA publishes it. It does not measure policy enforcement or approval decisions on customer agent actions.
Reading the live record.
AI Risk Disclosure
The MIT AI Risk Initiative studied how the world's largest companies disclose AI risk, and recommended that reports say which mitigation addresses which risk. This is ALEETH's disclosure in that form. For each of the 24 risks in the MIT AI Risk Repository: the ICA control that addresses it, or where the risk sits when no platform control can.
14 with an ICA control
3 set by people and policy
7 beyond any one organization
1.1 Unfair discrimination and misrepresentation
ICA checks governed outcomes against the EEOC four-fifths rule and the NYC Local Law 144 metrics when an assessment runs. For healthcare customers it also checks appeal outcomes by subgroup daily, and a breach opens a finding for human review.
ICA control1.2 Exposure to toxic content
A narrow detector flags hate, harassment and violent-threat phrasing in AI-written text.
ICA control1.3 Unequal performance across groups
The same fairness check compares outcome rates across groups in the data it is given.
ICA control2.1 Compromise of privacy
Pattern detectors for health information, and a data-loss guard on outbound payloads, flag or block what a governed agent sends out. Some are switched on per customer.
ICA control2.2 AI system security vulnerabilities and attacks
ICA decides every governed agent tool call and where it may connect, on hosts where the egress lock is on, and its fail-closed harnesses run on every ICA code merge. Attacks on the model itself are the model provider's to stop.
ICA control3.1 False or misleading information
Agents are scored for reliability before release. This tests whether an agent performs, not whether what it says is true.
ICA control3.2 Pollution of information ecosystem and loss of consensus reality
A societal or market-level risk that no control inside one organization addresses.
Beyond one organization4.1 Disinformation, surveillance, and influence at scale
Set by the people who deploy and use AI. ICA enforces the policy they write for governed agents and records the decisions they make.
People and policy4.2 Cyberattacks, weapon development or use, and mass harm
A content detector flags weapons, CBRN and malware phrasing in agent content. A sandbox gate, where switched on, holds code-execution calls that lack a signed sandbox attestation. An egress lock limits where a governed agent can reach.
ICA control4.3 Fraud, scams, and targeted manipulation
Set by the people who deploy and use AI. ICA enforces the policy they write for governed agents and records the decisions they make.
People and policy5.1 Overreliance and unsafe use
Set by the people who deploy and use AI. ICA enforces the policy they write for governed agents and records the decisions they make.
People and policy5.2 Loss of human agency and autonomy
A consequential agent action can require a named person's approval at the gateway, and an autonomy tier ladder sets how far each agent may act on its own.
ICA control6.1 Power centralization and unfair distribution of benefits
A societal or market-level risk that no control inside one organization addresses.
Beyond one organization6.2 Increased inequality and decline in employment quality
A societal or market-level risk that no control inside one organization addresses.
Beyond one organization6.3 Economic and cultural devaluation of human effort
A societal or market-level risk that no control inside one organization addresses.
Beyond one organization6.4 Competitive dynamics
A societal or market-level risk that no control inside one organization addresses.
Beyond one organization6.5 Governance failure
ICA watches the official regulatory sources it is set up for and maps each change it finds to the controls it affects, for human legal review.
ICA control6.6 Environmental harm
A societal or market-level risk that no control inside one organization addresses.
Beyond one organization7.1 AI pursuing its own goals in conflict with human goals or values
ICA compares each agent's tools and scopes with the set it was certified for, and its behaviour with its own baseline, and flags drift.
ICA control7.2 AI possessing dangerous capabilities
The gateway holds each governed agent to the tools and scope it is certified for, whatever the model behind it can do.
ICA control7.3 Lack of capability or robustness
Agents are scored for reliability before release, and a fail-closed test matrix runs on every ICA code merge: in armed mode, when a control cannot decide, the action is held or refused.
ICA control7.4 Lack of transparency or interpretability
Every governed decision is sealed as a signed receipt in a verifiable transparency log, and ICA publishes its own enforcement coverage in the live record above.
ICA control7.5 AI welfare and rights
A societal or market-level risk that no control inside one organization addresses.
Beyond one organization7.6 Multi-agent risks
ICA records which agent asked which agent to act, so a chain of delegated actions traces back to the request that started it.
ICA controlA mapping, not a certificate or a score. Some controls are switched on per customer. This is a snapshot, read from ICA's published map at commit abd70cba2 (sha256 b4e7549df445…), 2026-11-11. Risk ids and short titles from the MIT AI Risk Repository, CC BY 4.0, credit MIT AI Risk Initiative.
Platform Controls
01.
ICA writes control events to an Ed25519-signed, hash-chained ledger. Database triggers refuse updates and deletes on the ledger tables. On a recurring cadence the ledger's Merkle root is anchored outside ICA through OpenTimestamps, so an edit or deletion made after anchoring contradicts the anchored record.
02.
Connections to aleeth.com and to ICA use TLS, and plain HTTP is redirected to HTTPS. ICA reaches its database over HTTPS, and the database is encrypted at rest with AES-256. Anthropic and OpenAI Admin API keys that a customer connects through the provider-connection API, and single sign-on client secrets, are encrypted with AES-256-GCM before they are stored, and the write is refused if the encryption key is missing. That API returns only a masked fingerprint of a stored key, never the key.
03.
Every request is authorized against the tenant it targets in the API layer. A customer token scoped to one tenant receives a 403 when it reaches for another tenant's data, and ICA's own test harness checks this. ALEETH operator accounts can act across tenants. Row-level security policies at the database provide defense in depth for direct client access.
04.
Customer tokens are scoped to a single tenant by signed claim.
05.
State transitions record the actor, the before and after state, source address, and timestamp to an append-only audit log enforced by a database trigger, which the application cannot rewrite.
06.
An internal sentinel probes platform health and integrity on a fixed cadence and writes any incident to an append-only log, with alerting. We hold our own platform to the same continuous-evidence standard we require of the systems we govern.
Agent Governance
Securing the platform is the floor. These govern the agents themselves: their purpose traces to the business, and their fairness is measured.
Strategy & Vision
Every governed agent traces to a business objective, a named owner, and a risk tier, rolled up across the estate with an honest coverage posture. Uncharted agents are counted, never shown as passing.
Ethical Principles
Bias measured on the EEOC four-fifths rule and the NYC Local Law 144 metrics. Advisory by design, it flags for human review and never silently moves an agent's standing.
Data · Residency · Subprocessors
Primary data residency is the United States. Subprocessors are listed below by category. We do not sell customer data, and ICA does not train language models on customer evidence. ICA's detection tuning model learns from past findings, including their tenant and actor, and whether each was confirmed or dismissed; it proposes rule changes and never applies them.
Responsible Disclosure
We welcome good-faith security research. If you report a vulnerability in good faith and within the scope below, we will not pursue legal action, we will acknowledge your report, and we will keep you updated through remediation.
REPORT TO
security@or via the secure inquiry form
The full vulnerability disclosure policy
DO report promptly, give us reasonable time to remediate before public disclosure, and act in good faith.
DO NOT access, modify, or exfiltrate another customer's data, degrade or interrupt service, or run automated scanning that disrupts availability.
SCOPE aleeth.com and its public verification surfaces (the public credential verifier and trust pages). Operator and customer consoles, and social engineering of staff or customers, are out of scope.
In progress
What is not held
Next · Provenance
Every statement on this page about signed, append-only, anchored records is verifiable. The live receipt chain and the public-ledger anchor are open to any browser, with no account and no permission required.
See the live proof →