Safe by structure, not by prompt.
Telling an AI not to do something is a request. Leaving the tool out of reach is a guarantee. Here's what the AI can and can't do in Malleable, how your data is handled, and the controls IT gets.
Limits on the AI.
Each workflow runs in its own harness. The limits live there, where no prompt can talk its way past them.
Each stage uses only the tools it's given.
The AI can't reach a system the current stage doesn't need, and it can't add tools mid-run.
Fixed steps where it matters.
Your team chooses which steps run the same way every time and where the AI can use judgment. An input you lock, like an account or a recipient, is never the AI's to change.
People decide what's theirs.
A stage that needs a person waits for their decision before anything changes.
A stage finishes only when its work is done.
A stage can't close until it has collected the data it needs. Then the run can move only to a stage the workflow diagram connects to.
Every run is on the record.
Each run records what the AI did, who decided what, and the exact form each person saw.
Nothing goes live without a person.
Malleable can suggest fixes, but a person approves each one.
Malleable Bot asks first.
In Slack and Teams, Malleable Bot asks before it uses anyone's access or changes anything.
Code runs in a sandbox.
Any code Malleable writes runs in its own isolated sandbox, a new one for each run.
IT sets the limits. Teams build inside them.
Letting the people who run a process build it doesn't mean losing control. Teams build on what IT has approved, and every change is on the record.
Approved tools only
On Enterprise, admins choose which integrations and models teams can use.
Real access, not a super-user
Each workflow acts with the account of its owner, or of the person who started the run.
Each person sees their part
Each stage names who sees it, like the requester or the approver.
Every version is kept
Rollback is one step.
Your data
- Encryption
- Encrypted in transit with TLS and at rest.
- Model training
- Your data is never used to train AI models.
- Subprocessors
- We vet every provider that processes your data, and every one is listed in our Trust Center.
- Retention and deletion
- Ask, and we'll delete all of your organization's data. Enterprise customers can set how long runs are kept.
Sign-in, apps and outside people
- Secure login
- People sign in with single sign-on, or a short-lived code sent by email.
- Connected apps
- You connect each app through its own sign-in screen. We store a token, never your password, and you can revoke it at any time.
- Outbound requests
- Every request a workflow sends goes through a proxy that blocks private addresses.
- Outside people
- People without a Malleable account answer from a secure, unguessable link unique to that one request.
Our own security
- Production access
- Only staff who need it can reach our live systems, over encrypted connections with multi-factor authentication.
- Vulnerability management
- We scan our systems continuously, fix critical issues, and patch servers routinely.
- Independent audit
- Prescient Assurance, an independent auditor, tested our controls over several months and attested that they meet SOC 2 Type II.
- Report an issue
- Email security@usemalleable.com.
Questions from your security team?
We'll walk through the controls, the Trust Center, and how a first workflow would be set up.