Learn · Security

Read something untrusted? Ask before acting.

Taint tracking marks a run as tainted the moment it reads untrusted content. A tainted run can keep working, but its high-risk actions go to a human.

Start your free monthAsk your agent to try it

Mark

Taint follows the run

Reading an email, web page or another agent's message taints the run. Reading only your own directives doesn't.

  • Trust levels decide taint
  • Taint is shown on approval cards
  • Logged in the audit trail

Limit

High-risk actions downgrade

In a tainted run, actions like sending outside, deleting, schema changes and company key use become ask-owner, whatever the autonomy dial says.

  • Overrides Autopilot for risky actions
  • Low-risk work continues
  • No silent exfiltration

Decide

You see why it's asking

The approval card says the run was tainted and shows what it read, so you can judge.

  • Tainted badge
  • Source of the untrusted content
  • One-tap decision

Questions

Does taint tracking slow agents down?

Only for high-risk actions in runs that read untrusted content. Everything else carries on.

Keep exploring

The dual-LLM patternA quarantined model reads untrusted text; the privileged one never does.Prompt injection, for agentsHow injected text hijacks agents, and the defenses that work.One-tap approvalsRisky moves wait for your yes on your phone, email or Slack.

Taint tracking, on by default

Start your free month.

Start your free monthAsk your agent to try it