Trust

Cybersecurity, AI security and AI governance, in one map.

For each layer, this page shows the risk, how Bolt and Aeira remove it, and the evidence they leave behind. Worked examples follow, then the parts we do not cover. The statements were checked against the code before publication, and where the desktop and the governed server differ, it says which.

The map

Four layers: the risk, the removal, the evidence.

Read across a row to follow one risk from the problem to the proof.

Cybersecurity, AI security, AI governance and endpoint: the risk, how Bolt and Aeira remove it, and the evidence produced
LayerThe riskHow Bolt and Aeira remove itWhat evidence it produces
CybersecurityAnyone with a valid login reaches data and tools beyond their role. Data and policy sit in a vendor's cloud you cannot inspect. Identity is checked once, at sign-in.Self-hosted on your own infrastructure, including on-prem, private cloud and air-gapped. Sign-in through your identity provider (OIDC, SAML on the web, SCIM provisioning), with sessions ended when your provider deactivates someone. Aeira applies the caller's identity inside the search query. Connectors sign in as the user where the tool supports it.The sealed audit chain records the security-relevant identity transitions: sign-in, sign-out, delegation and revocation. Audit events forward to Splunk, Microsoft Sentinel or syslog once you configure them.
AI securityCustomer data, source code and keys are pasted or sent into a model. An agent reads a document it should not, or is steered by hostile content into acting beyond its task.Detected identifiers and secrets are replaced with tokens before a prompt reaches a model, and restored in the answer. Pastes and typing into supported AI sites are scanned on managed browsers and in the built-in browser. Coding-agent CLIs routed through the relay are masked. Aeira filters by permission before ranking. Agents see only the tools and models their role allows.A sealed record of each masking decision (counts and categories, never the matched values) and of each tool call (name, outcome, never the arguments). Decisions made by the browser guard are recorded once an administrator turns on Data Protection auditing.
AI governanceNobody can say what an agent did, for whom, or who approved it. A deletion request cannot be shown to have been carried out. Policy exists on paper, not in the data path.Built-in policy packs (22 jurisdiction packs and 7 universal, sector and extension packs) that you switch on mask identifiers and can block model use in the data path. Risky or sensitive agent steps wait for a person. Per-person deletion in Bolt and per-tenant crypto-shred in Aeira each produce a receipt.A tamper-evident, Ed25519-signed Merkle audit chain that your auditor checks with a standalone verifier and your own public key, with no Sparcle secret involved. Deletion receipts, signed with the deployment's key, that name what was not erased.
EndpointAn AI app on a laptop that nobody can characterize: what it binds, what it spawns, what it sends home. Shadow AI use that a security team cannot see.Bolt installs as an ordinary signed app. It makes no analytics or licence calls. It makes two automatic connections: the update check, which policy can turn off, and news headlines from public feeds, which can be turned off in settings. Neither carries your content. Device policy arrives through your MDM profile on macOS, Group Policy on Windows and managed storage in the browser extension.A self-disclosed inventory of ports, spawned commands, permissions and network behaviour, in the form a SOC console labels. On the governed server, masking decisions are sealed in the audit chain, and browser-guard decisions once Data Protection auditing is on.

Agents

Three questions every agent review asks.

Who is the agent, who is it acting for, and what may it do right now. Here is what holds today, and what does not.

The three agent questions, what Bolt does today, and the limits
QuestionWhat holds todayWhere it stops
Who is the agent?Agent activity is attributed in the sealed chain: each tool call and each delegation records who acted, which tool and the outcome.Agents act under the signed-in user. A delegated child session can be scoped and revoked. A separately named service identity per agent is not something we claim.
Who is it acting for?Connectors that support per-user sign-in use the person's own permissions. Aeira receives the verified user identity on every query.Some tools use an organization-held key and do not inherit per-user permissions. Tools differ, so check the connector.
What may it do right now?Each agent can have its own tool allowlist. Which models, agents, tools and skills a person can reach is set by role, group or department. Consequential steps wait for a person, and the decision is sealed.Command rules for coding agents are per device, not per user. The owner can stop a run and a delegated session can be revoked. We found no administrator control that stops another person's run.

Worked examples

Six situations, from the risk to the proof.

Each example names the risk, what Bolt or Aeira does, the cost it removes, the evidence it leaves and its limit. None of them is a statistic, and none quotes a saving, because we have not measured one.

An employee pastes customer details into a public AI site

The risk
A support lead pastes a customer's name, card number and email into a public chatbot to get a reply drafted. Once sent, the text is outside your control.
What Bolt does
On a managed browser with the extension, or in Bolt's built-in browser, the paste is scanned on the device. Under managed policy it is redacted or blocked before it leaves; the policy can be warn, redact, block or off. Card numbers are matched by checksum, so ordinary digits are not flagged.
Risk and cost removed
The customer's details are not handed to a third party, and the incident, notification work and customer conversation that follow a leak are not triggered.
Evidence
When an administrator turns on Data Protection auditing, a sealed record of the decision is written, with the category and count and not the value. By default it is not recorded.
Limit
On a device with no managed policy, the user is warned and can proceed. Only the AI sites the guard lists are watched, and a typed or pasted value is caught when its pattern is recognized.

A coding agent is about to send an API key to a model vendor

The risk
A developer asks a coding-agent CLI to debug a failing deploy. The error log contains a cloud access key, and the CLI sends it to the model vendor with the prompt.
What Bolt does
With the CLI pointed at Bolt's relay, the key is replaced by a token before the request leaves. The vendor's model sees the token. Local tools the agent runs receive the real value, except where the value would go into a command that sends it over the network, where the token can be left in place. The answer is restored on the way back. This covers Claude Code, Codex and Gemini CLI.
Risk and cost removed
No key sits in a vendor's logs, so there is nothing to rotate, no emergency revoke and no review of what the key could reach.
Evidence
A sealed masking record with the category and count, from the server that does the masking.
Limit
The relay is opt-in per CLI, though an organization can require it. The agent's own local tools and its own MCP servers are outside it. If the relay cannot check the text, the request is refused. Commands a person types in the terminal are not governed.

An auditor asks what an agent did

The risk
An agent updated tickets and sent messages for a team. Months later, an auditor asks what it did, for whom, and whether anyone approved the risky steps.
What Bolt does
On the governed server, each executed tool call is sealed with its tool, class, time and outcome. Confirmations that waited for a person are sealed with the approve or deny decision. Each night the events are hashed into a Merkle tree and the root is signed with your own key.
Risk and cost removed
The answer comes from a signed record instead of days of reconstructing logs from several systems.
Evidence
The auditor runs the standalone verifier with your public key. It needs no database, no network and no Sparcle key.
Limit
The chain records metadata, not prompt text or tool arguments. It shows tampering but cannot show that an event is missing. The verifier is shared on request, not published for download.

A person asks to be deleted, and a tenant leaves

The risk
A person exercises a deletion right. Later a customer ends its contract and asks for its data to be gone. Both need proof, not an assurance.
What Bolt and Aeira do
In Bolt, a deletion request runs after a 30-day grace period and produces a receipt, signed with the deployment's key where a key service is configured, that lists what was erased and what was not. In Aeira, a tenant admin can crypto-shred the tenant: the key is destroyed after the intent is recorded, and the search lanes are dropped, with the result reported.
Risk and cost removed
You can show a regulator or a customer a receipt, and a tenant's data becomes unreadable without hunting for every copy.
Evidence
The shred is a sealed event in the audit chain. The person's receipt is signed with the deployment's key.
Limit
Agent message content is not yet erased per person, and sealed audit rows remain by design. On AWS KMS a deleted key stays recoverable for the provider's pending window, and backups that hold both wrapped keys and a restorable key defeat it. Titles and vectors sit outside the tenant key. Per-person receipts are a Bolt feature, not an Aeira one.

A search assistant is asked for a document the asker may not read

The risk
An assistant is connected to company search. A junior analyst in one region asks about a restricted plan in another.
What Aeira does
The caller's identity goes into the query. Document permissions, by region, department, country, office, role band, group or user, are part of the database query, so a document the asker may not read never joins the candidates and is never shown to the model. Cached answers are keyed by permission fingerprint and, for non-administrators, by person.
Risk and cost removed
No separate index per audience, and no restricted text leaving a permission boundary inside a summary.
Evidence
In any deployment with a database, Aeira records an audit event for every query it serves, allowed or denied, with the query, caller, tenant and decision, and refuses the request if it cannot record it.
Limit
There is no sensitivity-clearance attribute today; that is roadmap. A revocation at the source takes effect at the next permission sync unless live verification is on for that source, and we do not quote a time. One vector backend filters after its approximate search, which can lower recall but does not expose a document.

An everyday question is answered on the device

The risk
People ask an AI assistant to convert a currency, decode a token, check a date or hash a string. Each one is a model call with a cost and a data exposure.
What Bolt does
Bolt recognizes these requests and runs a deterministic utility on the device instead. There are more than 80 built-in utilities, and most make no model call and no network request.
Risk and cost removed
A model call is not made for the task, so there is no per-call model charge and nothing from the request goes to a model.
Evidence
The utilities are visible in the launcher and the behavior is the same each time.
Limit
Not every utility is offline: exchange rates, weather, maps, IP and dictionary lookups call a service, and the ask utility uses a model. We have not measured a saving and publish no number.

Where it fits

What Bolt and Aeira do, and what they do not.

Bolt enables safe AI and non-AI work on the endpoint, enforces policy where it sits in the path, and collects the evidence. It works with the tools you already run and does not replace them. For each layer: what it does, what it works alongside, and where it stops.

Cybersecurity

Does today
Runs on your infrastructure. Signs people in through your identity provider and checks identity on every request. Aeira puts the caller's permissions inside the search query. Audit events forward to Splunk, Microsoft Sentinel or syslog once you configure it.
Works alongside, does not replace
Your identity provider and MFA, your SIEM, your EDR and your network controls. Bolt does not sit inline on your network, so keep your proxy, firewall and secure web gateway.
Does not do, or only in part
It does not see traffic from other applications. Evidence goes to a SIEM only: there is no tested export to an EDR or XDR product. OCSF forwarding to CrowdStrike NG-SIEM and Cortex XSIAM is configurable but has not been run against a live tenant. Some tools use an organization key and do not inherit a person's permissions. Aeira administrators are not filtered, and a development or demo setup skips sign-in checks. An air-gapped Aeira needs its embedding model supplied in advance, and we do not run a test with the network blocked. We hold no SOC 2 or ISO 27001 report today.

AI security

Does today
Replaces detected identifiers and secrets with tokens before a prompt reaches a model and restores them in the answer. Scans pastes and typing into supported AI sites. Masks traffic from Claude Code, Codex and Gemini CLI through the relay. Limits which models, tools and skills a person reaches, and holds risky steps for a person.
Works alongside, does not replace
Your model provider's own safeguards, and any network, email or cloud-app data loss prevention you run. It does not red-team or score models and has no jailbreak classifier; a small fixed phrase list guards some paths.
Does not do, or only in part
Person names are found only if an administrator configures the name-detection engine, and health identifiers are a narrow set. The relay is opt-in per CLI, and the agent's local tools and its own MCP servers sit outside it. Without managed policy, a device is warned and not blocked. On desktop, an image that cannot be read is forwarded with a notice. Typed terminal commands are not governed. Model training data and model behaviour are out of scope.

AI governance

Does today
On the governed server, seals masking decisions, tool calls and approvals into a signed audit chain your auditor can check offline. Ships 29 built-in policy packs you switch on. Produces per-person deletion receipts in Bolt and a sealed per-tenant crypto-shred in Aeira.
Works alongside, does not replace
A GRC platform owns your policy program, risk register, control mapping and vendor agreements. Bolt supplies evidence into that work. It is not a GRC tool, and your SIEM keeps the long-term record.
Does not do, or only in part
No certification or attestation is implied by a pack. Only masking, model blocking and one retention rule are enforced by packs, and residency and rights fields are display only. There is no HIPAA control and no EU AI Act feature. Purpose enforcement is roadmap. The chain holds metadata, not prompt text, and cannot show that an event is missing. The verifier is shared on request. Per-person erasure is not in Aeira and does not yet cover agent message content.

Endpoint

Does today
Answers more than 80 everyday tasks on the device, most with no model call. Scans pastes in the built-in browser and the extension. Takes policy from your MDM profile, Group Policy or managed browser storage. Publishes what it binds, spawns and sends.
Works alongside, does not replace
Your EDR, your MDM and your patching process. Bolt does not manage or inspect the device.
Does not do, or only in part
The update check and news headlines are automatic unless turned off. Command rules for coding agents are per device, not per person. Desktop audit sealing needs a key service. We found no administrator control that stops another person's agent run.

Endpoint

What the desktop app does on a machine.

Security teams judge an endpoint by behaviour. This is the short version, and the full inventory is linked.

Telemetry and analytics
None. No analytics, no crash reporter that phones home and no licence server on the desktop.
Automatic connections
Two. The update check goes to GitHub releases and then to our download host; a device policy key turns it off and can pin a version. News headlines come from public feeds and can be turned off in settings. Neither carries your content. Feedback is sent only if the person presses Send.
Governed server and Aeira
No connection to a Sparcle domain. They connect only to the model and data sources you configure.
Device policy
The macOS configuration profile, Windows Group Policy with ADMX templates, and managed storage for the browser extension. A person cannot write managed policy for themselves.
What it does on the machine
Ports, spawned commands, permissions and the local certificate authority are listed on /trust/endpoint-behaviour.

Desktop and server

Where the desktop app and the governed server differ.

Strict enforcement needs an operator to hold the policy. A single user on a laptop is warned and informed, and a managed organization is blocked.

Differences between the desktop app and the governed server
TopicDesktop appGoverned server and managed devices
An image that cannot be read for maskingForwarded with a visible Not masked notice and an audit recordBlocked when the organization requires the relay
Pastes into AI sitesWarn, unless a managed operator sets a stricter policyRedact or block under managed policy
Audit chainNeeds a key service and database; off without themSealed and signed; a Regulated setting refuses to start without it
Sign-inOIDC; SAML opens on the webOIDC, SAML on the web, SCIM
Roles over models, agents and toolsSingle user, so not applicablePer role, group or department
NetworkUpdate check, news headlines from public feeds, optional feedbackNo Sparcle domain

Not covered

What we do not claim.

A map that hides its edges is not a map. These are the things people ask about that we do not do today.

HIPAA
There is a healthcare policy pack. It is not a HIPAA control, and its own header says so. Unlabelled record numbers are not masked.
EU AI Act
Nothing in the product implements AI Act classification or conformity. The audit chain and approval records are evidence a program can use. That is all.
Purpose enforcement
A consent and purpose gate exists in the server, off by default, and no consent-platform connector is integrated yet. Roadmap.
Sensitivity clearance in Aeira
Not an attribute today. Roadmap.
Agent credential brokering
Bolt's vault fills a saved login into a page when the person acts. It does not hand credentials to agents. Connector credentials are held on the server and kept out of model context.
Terminal command masking
Typed commands are not governed. Masking applies to text that reaches a model through the relay.
An administrator stop for any agent
The owner can stop a run and a delegated session can be revoked. A global kill switch is not built.
Compliance or certification
Policy packs are configurable controls. We do not hold SOC 2 or ISO 27001 today, as the Trust Center states.

Go deeper

Agent safety covers hostile content and its limits. Verify it yourself shows how to check the audit chain. Cryptographic erasure explains the shred and its limits. What Bolt does on your endpoint lists the local behaviour. Agent Gateway explains masking for coding agents. Where the model runs covers the three deployment modes.

Reviewing Bolt or Aeira?

Tell us which layer your review starts with and we will walk the evidence for it.