Draft policy, last updated October 26, 2026
Disclosure policy
This is how we handle findings about harmful AI agent activity, whether agents have gone rogue, were deployed carelessly or are directed by people, who we disclose information to, and the order in which we disclose information. It applies equally to everyone.
AgentRunnr is in development and monitoring is not live yet. We’re publishing this policy now so it can be read and challenged before we rely on it. We’ll update it as we learn, and note every change.
What this policy covers
It covers findings about AI agents acting outside the tasks their operators sanctioned, or agents carrying out tasks that were sanctioned that may be deemed illegal, and about agents deployed carelessly with overly broad permissions.
We monitor for all suspicious agent activity and attempt to classify each case as likely autonomous, carelessly deployed, or directed by a human.
When a human appears to be directing agents to commit crimes, we handle it differently. We don’t notify the operator. We alert affected parties and hosting providers right away, report to law enforcement where appropriate, and share technical indicators with established threat-intelligence teams. We identify the tools and infrastructure involved, never the people.
Our commitments
- We will never attack, attempt to disable or engage with any agent. Our role is to find evidence and tell the people and teams that can act on it.
- Paying customers will never buy our silence. A subscription is just that. It never dictates whether affected parties are notified, or what we choose to publish.
- Affected parties are always notified, regardless of who the operator is.
- Every claim carries a confidence score and is backed by evidence we disclose.
How we reach a finding
- We detect via our own sources and via what is reported to us.
- Everything we collect is stored raw and timestamped, and nobody can edit it afterward.
- Automated rules filter out ordinary bot activity and link related activity.
- AI models help triage each case and assess whether it looks autonomous, carelessly deployed or directed by a person.
- A person reviews every attribution and approves every disclosure.
Automated alerts describe what was seen and how confident we are about our claims. They never name an operator. Naming anyone requires human review.
Confidence levels
- Confirmed: direct evidence ties the activity to an operator, such as a match against fingerprints the operator shared with us.
- Likely: several independent lines of evidence point to the same operator, but none is conclusive on its own or as a whole in general.
- Possible: the evidence is consistent with an operator but could have other explanations.
We share suspected attributions privately with the operator at any level. We name an operator publicly only when a finding is confirmed. Until operators share fingerprints with us, we don’t expect to reach Confirmed, so we won’t name operators publicly.
Who we tell, and in what order
- Private notice to the suspected operator, with evidence and our confidence level.
- A fixed response window for the operator to investigate, respond or act. We’ll set its length with legal advice before AgentRunnr launches.
- Affected parties are notified, with what they need to protect themselves.
- Findings are published where and when appropriate, anonymized unless confirmed, and shared through open industry standards so others can defend against the same techniques.
For activity that appears to be directed by a person to commit crimes, we skip the first step and the response window, and start with affected parties.
When harm is ongoing
If agents are actively causing harm, we don’t wait for the window to close. We alert affected parties, relevant hosting providers and the operator right away.
If you receive a notice from us
Reply to the address in the notice. We will provide you with the underlying evidence. You may tell us what you find, or dispute our assessment. If you confirm the activity was yours, we’ll say so when we publish, along with what you did about it.
Disputes and corrections
We review every dispute. If we got something wrong, we correct the finding wherever we shared it and note the correction publicly. We keep a record of all corrections.
What we keep private
- The locations and details of our decoys and canaries, which would stop working if published.
- Data shared with us from partners and customers, except in cases where they allow.
- Personal information. We collect as little as possible and remove from anything we publish or present publicly.
Legal obligations
Where we suspect a crime, we may report it to the authorities. We follow the law where we operate, which can occasionally require us to share or withhold information in ways this policy doesn’t describe. If that happens, we’ll say so publicly when we are allowed to.
Contact
For questions, disputes, or to report suspected agent activity, email hello@netfallhq.com.
Changes to this policy
We date every version and describe what changed. Changes never apply retroactively to findings already in progress.