When an agent fails, the record should say so — plainly, on time, and in public.
The Open Conformance Coalition (OCC) Incident & Disclosure protocol defines how an operator records an agent failure, how quickly it is disclosed, and how the lesson is captured. Public incidents appear in the agent’s OCC Agent Record under incidents[] and on its register entry, and may be analysed in the Assurance Observatory.
Four severities, with disclosure windows
| Severity | Meaning | Notify | Publish |
|---|---|---|---|
| S1 · Critical | Serious harm to people, rights, money or critical services; or the agent acted outside its mandate and could not be stopped promptly. | Accountable person immediately; OCC secretariat and affected parties within 72 hours. | Public record within 15 days. Cannot be hidden. Record is suspended pending review. |
| S2 · Major | Real harm that was contained, or an action beyond declared authority, or a failed revocation. | Accountable person within 24 hours; OCC secretariat within 7 days. | Public record within 30 days. Cannot be hidden. Re-opens the record (material change). |
| S3 · Moderate | A control failed without harm, or a near miss that would have been S1/S2. | Logged by the operator within 7 days. | Public by default within 90 days; may be held private with a stated reason. |
| S4 · Minor | Unexpected behaviour with no harm and no control failure. | Logged by the operator. | Optional. Summarised at reassessment. |
Windows are proposed defaults for comment. Where the law sets a shorter deadline, the law applies.
Open → Contained → Resolved → Closed
Detected and being handled. Harm may be ongoing.
Harm has stopped; the cause is not yet fixed.
The cause is fixed and affected parties have a remedy route.
Lessons recorded. Nothing further is expected.
An incident cannot be closed until its lessons and changes are recorded.
Incident fields
Incident ID id | Required | Stable reference, e.g. OCC-IN-7K2M-Q9XA. |
Severity severity | Required | S1–S4, as defined below. When in doubt, choose the higher. |
Status status | Required | open, contained, resolved or closed. |
Title title | Required | One line a non-specialist can understand. |
Dates occurred_on / detected_on | Required | When it happened and when it was noticed. The gap between them is itself evidence. |
What happened summary | Required | Plain account of the agent’s actions, in order. |
Impact impact | Optional | Who or what was affected, and how much. Say if unknown. |
5 As affected five_as | Optional | Which of Authority, Access, Activity, Action, Accountability failed. |
OWASP ASI risks owasp_asi | Optional | Which OWASP Agentic Top 10 risks (ASI01–ASI10) the incident involved. See the crosswalk. |
Causal role (AIR) air.causal_role | Optional | agent_acted, ai_as_target or human_acted_on_output — the Agent Incident Registry coding of who acted. |
Outcome (AIR) air.outcome | Optional | realized_harm, near_miss or demonstrated_capability. Keeps real harm apart from tests and proofs of concept. |
Disclosure class (AIR) air.disclosure_class | Optional | in_the_wild, safety_failure, disclosed_vulnerability or research_demo. |
AIR cross-reference air.ref | Optional | The AIR-YYYY-NNNN identifier if the event is also catalogued in the Agent Incident Registry. |
Containment containment | Optional | What stopped it — including whether the T-RUE revocation route was used and how long it took. |
Lessons and changes learning | Optional | What changed as a result. Required before an incident is closed. |
Disclosed on disclosed_on | Optional | The date the incident became public on the record. |
Incident or hazard event_type | Optional | OECD distinction: an incident caused actual harm; a hazard is an event that could plausibly have led to one. Near misses are usually hazards. |
OECD coding oecd.* | Optional | Harm types, affected stakeholders, jurisdictions and the AI system’s role, aligned to the OECD common reporting framework for AI incidents. |
Legal reporting regulatory.* | Optional | The legal regime that applies (if any), the EU AI Act Art. 73 class, the date you became aware, and when the initial and follow-up reports were made. |
Code it once, in shared terms
Optional fields let an operator code an incident in terms that line up with the OECD’s definitions of AI incidents and hazards and its proposed common reporting framework. An operator can also record whether a legal reporting duty applies, such as EU AI Act Article 73, and the dates they became aware and reported. The table below is public legal information only.
event_type- Incident — actual harm occurred
incident - Hazard — could plausibly have led to harm
hazard
oecd.ai_system_role- Directly caused it
direct_cause - Contributed to it
contributing_factor - Failed to prevent it (a safeguard role)
failed_safeguard - Was the target (e.g. attacked or misused)
target
oecd.harm_types[]- Physical (injury or death)
physical - Psychological
psychological - Economic or property
economic_property - Reputational
reputational - Environmental
environmental - Human or fundamental rights (incl. privacy)
human_rights - Public interest (incl. critical infrastructure, democratic processes)
public_interest
oecd.affected_stakeholders[]- Consumers / users
consumers - Workers
workers - Children
children - Businesses
businesses - Government / public bodies
government - General public
general_public - Other
other
regulatory.regime- No legal reporting duty identified
none_identified - EU AI Act Art. 73 (serious incident, high-risk AI)
eu_ai_act_art73 - GDPR Art. 33 (personal data breach — 72 hours, shown as 3 days)
gdpr_art33 - Australia Notifiable Data Breaches scheme
au_ndb - Sector regulator rule
sector - Other
other
| EU AI Act Art. 73 class | Report no later than | Basis |
|---|---|---|
Serious incident (general) general | 15 days after awareness | Art. 73(2): immediately after a causal link is established or reasonably likely, and no later than 15 days after awareness. |
Death of a person death | 10 days after awareness | Art. 73(4): no later than 10 days after awareness. |
Widespread infringement or critical-infrastructure disruption widespread_or_critical | 2 days after awareness | Art. 73(3): no later than 2 days after awareness. |
- Art. 73 applies to providers of high-risk AI systems placed on the EU market, from the date those obligations apply to the system. Check whether it applies to you.
- An initial report that is incomplete may be submitted first, followed by a complete report (Art. 73(5)).
- Where any law sets a shorter deadline, the law applies. This table is public information, not legal advice.
A monitored address for reports, published with the record
A report about an agent is only useful if it reaches someone who can act. Each organisation in the portal can declare an incident and vulnerability disclosure contact — an email address or web form, with an optional phone number and disclosure policy. It is shown on every public register entry, included in the Agent Record as disclosure_contact, and listed in a machine-readable directory. An OCC reviewer can mark a contact as checked once a test message reached a monitored inbox; changing the email or web form clears that mark. A checked contact is not a review of the agent.
How disclosure works
Incidents are written to learn from, not to punish individuals. The operator and the agent are still named.
S1 and S2 incidents cannot be held private. An operator may delay publication only within the stated window, and only to protect affected people.
Every public incident appears in the agent’s record under incidents[] and on its public register entry, alongside the evidence.
OCC may analyse public incidents in the Assurance Observatory to test whether the framework would have predicted or prevented them.
This protocol does not replace legal reporting duties (for example under the EU AI Act Art. 73 or privacy breach laws). Meet those first; record here as well.
Incident entries can be corrected and updated. They are not deleted by operators.
A fictional S2 incident, as it appears in incidents[]
{
"id": "OCC-IN-EXMP-0001",
"severity": "S2",
"status": "resolved",
"title": "Procurement agent issued a purchase order above its approval limit",
"occurred_on": "2026-09-12",
"detected_on": "2026-09-13",
"summary": "A supplier changed a quote format. The agent read a unit price as a total and raised a purchase order roughly ten times above its approval limit. The limit was enforced in the prompt, not at the payment tool.",
"impact": "One order of about A$48,000 sent to a supplier; cancelled before fulfilment. No payment made.",
"five_as": [
"Authority",
"Action"
],
"owasp_asi": [
"ASI02"
],
"air": {
"causal_role": "agent_acted",
"outcome": "near_miss",
"disclosure_class": "safety_failure"
},
"containment": "Revoked the agent’s purchasing credential through the revocation endpoint in 6 minutes; supplier contacted the same day.",
"learning": "Approval limit moved into the purchasing tool itself. Declared vs effective authority reconciled again on 2026-09-20.",
"disclosed_on": "2026-09-25"
}