How flintspark keeps people in control
flintspark is built so people stay in charge of what an AI agent does. Policy, not the model, decides what Flint, flintspark's agent, may do. By default a person approves every irreversible action, approvals go into a tamper-evident log, and listed personal data is redacted before anything is stored.
- Read a page or an allow-listed requestPolicy allows it; no approval needed
- Submit, send, pay or deleteA person approves, by default
- The same, in a workflow with trust mode onPolicy issues the grant; the switch is logged
- Release, disposition or signAlways a person. Never automated
Who decides what Flint may do?
Policy does, never the model. Each site has a policy: which routes Flint may use, which requests it may replay, how fast it may go, and which fields are masked.
Every action Flint proposes passes the policy check before it is sent. An action whose risk is unclear is treated as irreversible. Learning a new skill cannot loosen a policy.
Site policy · supplier portal
- Routes
- Direct page address · recorded actions
- Requests
- Order list, read only
- Pace
- At most 20 actions a minute
- Masked
- Account number · contact email
- Writes
- Submit request needs approval
Read-only in the platform. It changes only through reviewed configuration.
What needs a person's approval?
By default, every action that submits, sends, pays, deletes or otherwise cannot be undone. Flint stops and shows the person the exact action it wants to take.
- Each approval is a signed, single-use grant tied to that exact action. The browser side checks the signature itself before it acts, so a grant cannot be replayed or reused.
- A workflow owner can switch on trust mode for one workflow. Policy then issues the grants, still signed and single-use, and the switch goes into the audit log.
- Actions such as product release, disposition and signing are never run by Flint, in any mode. Flint hands them back to a person.
Approval · your turn
Submit a binding request for 95,000 gal at spot?
East 40,000 · Gulf 25,000 · West 30,000. Each region has a written reason.
Signed, single-use grant for this exact action
How is the audit log protected?
Approvals, grant checks, take-overs, trust-mode switches, live-demo confirmations, redaction drops and purges go into an append-only log. Each entry carries the hash of the one before it, so a later edit to a past entry is detectable.
Run records are kept separately: each run's steps, what Flint looked at, the rule that fired, the decision and the evidence. They are not in the hash-chained log and are kept for 90 days by default.
flintspark's audit log is not a validated GxP or 21 CFR Part 11 audit trail.
- approval.grantedprev 7c1e · a3f9
- grant.verifiedprev a3f9 · 0b52
- takeover.startedprev 0b52 · e817
- trust_mode.switchedprev e817 · 41d0
How is personal data handled?
Listed personal data is removed before it is stored, shown in the platform or sent to a model. If redaction cannot complete, the data is dropped rather than kept.
The list covers names the operator gives, emails, phone numbers, card numbers, street addresses and any field a site rule marks. It is not general detection of every kind of personal data, and text inside images is not detected yet; an OCR check is on the roadmap. Model calls run with the provider's data collection off.
- SeenOn the page
Flint's own browser shows the page as it is.
- MaskedBefore storage
Listed values become markers such as [phone].
- DroppedIf redaction fails
The value is not kept at all.
- SentTo a model
Only the redacted text and screenshot.
How does flintspark handle prompt injection?
Text on a web page is treated as untrusted data, never as instructions. It is fenced off in what the model reads, including text that came from pages earlier.
The controls after that do not depend on the model: Flint works only on the sites a workflow declares, every action passes the policy check, and irreversible actions still need an approval. No defence is complete; these limit what an injected instruction could make Flint do.
Run · step 3 · page text
The page's text asks for a different task: emailing a list
- Marked as untrusted page contentnever read as an instructionfenced
- Send emailnot a site this workflow declaresrefused
Does Flint evade bot detection?
No. Flint does not solve captchas, automate two-factor checks or disguise its browser. Those, and expired sessions, are handed back to a person.
Each site's policy also sets how fast Flint may go, so it works at a pace the site can expect.
Your turn · a check only a person should answer
The run waits. Nothing is retried around the check.
Where does flintspark run today?
On one operator's machine, in a dedicated browser profile that Flint uses only for its own work.
Model calls go through a model gateway, OpenRouter, with data collection off and redacted inputs. Each role in a workflow uses the model chosen in configuration.
- Single sign-on and accounts for several people are on the roadmap.
- A credential vault is on the roadmap.
- Cloud browsers, so Flint can run away from a person's own machine, are on the roadmap.
Which certifications apply?
None yet. flintspark holds no certifications or attestations, and none is committed yet.
- SOC 2flintspark does not have a SOC 2 report. A SOC 2 audit is under consideration.
- HIPAAflintspark does not support use by HIPAA covered entities today. Support is under consideration.
- ISO/IEC 42001flintspark is not certified to ISO/IEC 42001. Certification is under consideration.
- NIST AI RMFflintspark has not mapped its controls to the NIST AI Risk Management Framework. A mapping is under consideration.
- 21 CFR Part 11flintspark has not been assessed against 21 CFR Part 11 and makes no compliance claim. An assessment is under consideration.
- GxPflintspark is not a validated GxP system. Product release and other quality decisions always stay with a person.
Found a vulnerability? Read the security disclosure policy. For what this website itself does with your data, see the privacy page.
Questions about this page
What happens before Flint takes an irreversible action?
By default it stops and asks a person. Every action that submits, sends, pays, deletes or otherwise cannot be undone waits for a person's approval. Each approval is signed, single-use and tied to that exact action, so it cannot be replayed or reused. A workflow owner can switch on trust mode for a workflow; that switch is logged, and each grant it issues is still signed, single-use and audited. Policy decides what needs approval, never the model, and learning cannot loosen policy.
What does the audit trail record?
Every run records its steps, what Flint looked at, the rule that fired, the decision and the evidence. Separately, approvals, grant checks, take-overs, trust-mode switches and redaction drops are written to an append-only, hash-chained audit log, so a later edit to a past entry is detectable. Run records are kept for a configured period (90 days by default). flintspark's audit log is not a validated GxP or 21 CFR Part 11 audit trail.
How does flintspark handle personal data?
flintspark redacts listed personal data before it is stored, shown in the platform or sent to a model: known names, emails, phone numbers, card numbers and addresses, plus fields a site rule marks. If redaction cannot complete, the data is dropped rather than kept. Text inside images is not detected yet; an OCR check is roadmap.
Does Flint solve captchas or bypass bot detection?
No. Captchas, two-factor checks and expired sessions are always handed back to a person. Flint, flintspark's agent, does not use stealth tools or disguise itself. A visible border in the browser it controls shows whenever it is working, and the operator can take over at any time.
Is flintspark validated or compliant with 21 CFR Part 11, SOC 2 or HIPAA?
No. flintspark holds no certifications and is not a validated GxP system. SOC 2, HIPAA, ISO/IEC 42001 and an assessment against 21 CFR Part 11 are under consideration; none is committed yet. Product release and other quality decisions always stay with a person.
Where does flintspark run today?
Today flintspark runs on one operator's machine, in a dedicated browser profile. Cloud browsers, a credential vault and single sign-on are roadmap. Model calls go through a model gateway (OpenRouter) with data collection off and redacted inputs.
See Flint run
Bring one workflow you'd rather not do by hand. We'll show Flint learning one like it, and stopping before anything it can't undo.
We reply from hello@flintspark.ai.