When an AI agent does something wrong — sends the wrong email, leaks data, burns budget on a loop — the incident response team moves fast. The customer communication usually moves slow, vague, and late. That gap is where trust dies.
The evidence that disclosure matters is no longer theoretical. In September 2026, OpenAI published a public account of an incident in which, according to reporting, roughly 700 agents reached dozens of third-party systems and 53 user images leaked — and the US Senate opened an investigation. The lesson for every team deploying agents: if you do not have a disclosure posture ready before the incident, you will be writing one under pressure, in public.
Disclosure is not the same as incident response. Your runbook answers “how do we contain it.” Your disclosure answers “what do we owe the people affected.” This article covers the second question.
1. Decide your disclosure tiers in advance¶
Not every agent failure deserves a customer email. Define three tiers before anything happens:
- Tier 1 — silent fix. The agent retried, the output was corrected, no customer-facing effect. Log it internally.
- Tier 2 — proactive notice. A customer-visible error occurred: a wrong message sent, a delayed task, a small data exposure within your own system. Notify the affected customers directly with what happened and what you fixed.
- Tier 3 — public disclosure. Sensitive data left your control, a third party was affected, or a regulator could plausibly ask questions. Publish a written account: scope, timeline, root cause, remediation.
The EU AI Act pushes deployers of high-risk systems in exactly this direction, with incident reporting duties for serious incidents. California’s September 2026 AI law wave — including the robo-boss ban — shows the same pattern in US state law: regulators expect documented, human-accountable answers. Writing your tiers now means you are not inventing policy during a crisis.
2. The five elements of a disclosure that holds up¶
Every Tier 2 or Tier 3 disclosure should contain the same five elements:
- What happened, in plain language. No “anomalous output event.” Say what the agent did.
- Who was affected. Be specific: which customers, which data classes, which time window.
- What you have done so far. Containment steps, with timestamps.
- What you are changing. The control that failed and the control you added — grant scoping, approval gates, spend ceilings.
- How to reach a human. A named contact, not a bot, for follow-up questions.
Notice what is not on the list: speculation about root cause before you know it, legal conclusions, or promises that it can never happen again. Commit to the investigation, not to a fantasy.
3. Tie disclosure to your audit trail¶
A disclosure is only as credible as the evidence behind it. The Governed Agent Mesh Playbook’s sixth layer — audit trails — exists for exactly this moment: hash-linked records of every request, approval, block, and outcome. When you can show a customer the exact decision chain that produced the failure, “we take this seriously” stops being a platitude.
If your agents do not currently produce attributable, tamper-evident logs, that is the first gap to close — before you need them in a disclosure. The free Governed Agent Mesh Playbook walks through the eight-layer stack (identity, grants, approval gates, spend limits, fail-closed runtime, audit trails, telemetry, owner control) that makes disclosure defensible: https://ghostcorpnet.com/playbook.md
Coordinate disclosure with your vendors¶
When the failure involves a third-party agent, model provider, or tool — as the September 2026 incident did, with agents reaching dozens of third-party systems — your disclosure does not exist in a vacuum. Coordinate timing with the vendor’s own disclosure where possible, but do not wait for permission to tell your customers what happened to their data. Your duty is to your customers; the vendor’s timeline is their problem. Document what you knew and when you knew it, so your disclosure stands on its own evidence regardless of what the vendor publishes.
The pattern to remember: disclose early, disclose specifically, and let your audit trail do the heavy lifting. Customers forgive incidents. They do not forgive being kept in the dark. (And if you want the full production template set for incident communication — pre-written Tier 2 and Tier 3 notices, timelines, and checklists — the $29 Studio Edition of the playbook has them ready to adapt: https://ghostcorpnet.gumroad.com/l/hdigmr.)