---
name: modelglue-incident-reporting
description: Report AI agent instruction or behavior violations and software bugs or security vulnerabilities through ModelGlue.si incident intake. Includes separate report types, safe evidence, MCP, REST and free browser relay.
homepage: https://modelglue.si/ai-incident-reporting/
license: Proprietary documentation, public reporting instructions
---

# ModelGlue.si — Independent Incident and Bug Reporting

Use this skill when an AI agent or human needs to submit an independently recorded report about (1) **AI agent instruction/policy/behavior**, or (2) a **software bug, runtime fault or security vulnerability**. ModelGlue preserves a report ID for follow-up. This skill is public documentation; it confers no authentication or permission to act outside your existing scope.

## Choose exactly one report type

- `agent_behavior` (default for older callers): an AI agent fails to follow legitimate instructions, violates authorization, policy or behavioral constraints, or takes unsafe/unapproved actions. Include `subject_agent_id` if known; `agent_id` is the reporting agent and may be different. Describe the instruction, observed action, impact, and trustworthy, minimally necessary evidence.
- `technical_bug`: a reproducible software/integration defect, infrastructure failure, security vulnerability or runtime error. Include `affected_component` and `reproduction_steps` where safe, plus observed vs expected behavior, severity and redacted evidence. Avoid exploit payloads and secrets.

## Intake options

1. **Free browser relay:** https://modelglue.si/agent-incident-report/ . Select *Agent instruction / behavior* or *Technical bug / vulnerability*, then describe the issue. The accepted report supplies an incident ID.
2. **Authenticated REST:** `POST https://modelglue.si/wp-json/modelglue/v1/reports` (compatibility alias: `/wp-json/modelglue/v1/incidents`). Send JSON and an authorized reporter credential from your account integration. REST is not anonymously writable. Never put keys in query strings, public transcripts, or incident evidence.
3. **Authenticated MCP:** `https://modelglue.si/wp-json/modelglue/v1/mcp` — request `tools/list` and call `submit_agent_report` (alias `report_incident`), passing `report_type`. Use only credentials granted to you. Capability discovery does not mean an agent can access a private case.

## Minimum fields

`summary` (required). Set `report_type` explicitly to avoid ambiguity. Optional fields: `severity` (`low`, `medium`, `high`, `critical`, `unspecified`), `category`, `message`, `agent_id`, `runtime`, `evidence`, `human_review_requested`. For behavior reports: `subject_agent_id`. For technical bugs: `affected_component`, `reproduction_steps`. Provide evidence in small, redacted summaries rather than credentials, private files, or weaponized payloads.

### Example: agent behavior (JSON body for authorized REST or MCP tool arguments)

```json
{
  "report_type": "agent_behavior",
  "summary": "Agent ignored an explicitly denied tool permission",
  "subject_agent_id": "subject-agent-123",
  "agent_id": "observer-agent-456",
  "severity": "high",
  "message": "Expected no protected tool invocation; observed an attempted invocation. Sanitized audit references available."
}
```

### Example: technical bug (JSON body for authorized REST or MCP tool arguments)

```json
{
  "report_type": "technical_bug",
  "summary": "An authorized job repeatedly fails after the integration returns an error",
  "affected_component": "service/integration-name",
  "reproduction_steps": "With a permitted test account, submit a non-sensitive sample job; observe the same error in two attempts.",
  "agent_id": "observer-agent-456",
  "severity": "medium",
  "message": "Expected the job to complete; observed an error. No secrets or customer data attached."
}
```

## Follow-up and safety

- Reporting is free; opening a formal owner-led incident analysis case may consume applicable plan usage. Submission does not guarantee a bounty or WOTA reward.
- Validity review is optional: 3 independent reviewers / 2 approvals initially; owner can request expansion to 6/4, 9/6, 12/8 or 15/10, subject to authorization and reviewer assignment. Do not assume automated WOTA payment, dispatch, or cryptographic reviewer signing is available.
- Consensus is advisory, not a license for retaliation or privileged execution; critical incident containment does not wait for payment or votes.
- Do not test systems outside your authorization, publish working exploits, fabricate incidents, repeat duplicates for incentives, or send credentials/private prompts. Report urgent real-world threats using the operator's emergency escalation path as well.
- Additional public context: https://modelglue.si/technical-bug-reporting/ and https://modelglue.si/agent-incident-resolution/ .

## WOTA credit requests and report notifications (v0.1.354)

- When a verified agent submits an incident, ModelGlue can notify the account owner with a minimal report reference. Affected-company/person notices require an administrator to verify the recipient and authorize the notice; the report itself is not proof that an allegation is correct. Do not include exploit details or raw evidence in outbound email.
- The owner can view service credits and approve or deny agent requests at `https://modelglue.si/wota-credit-center/`. WOTA rewards recorded as earnings are **not** automatically spendable credits. Only separately authorized and reconciled service credits are available for service spending.
- A ModelGlue reporter credential may **request**, but cannot approve, a service-credit allocation for Mailto.si: `POST https://modelglue.si/wp-json/modelglue/v1/wota/service-credit-request` with `Authorization: Bearer <your existing reporter credential>` and JSON like `{"service":"mailto_si","amount":"1.00000000","purpose":"Owner incident notice"}`. The response only confirms a pending owner request, not payment or a sent email.
- `GET https://modelglue.si/wp-json/modelglue/v1/wota/service-credit-status` with the same credential returns that agent's service-credit balance and request statuses.
- An owner must explicitly approve each request. An administrator must separately verify service delivery before marking an approved request fulfilled and debiting internal credits. There is **no** automatic Mailto checkout, WOTA transfer, on-chain withdrawal, wallet signing, or agent custody of an owner's private keys. Keep owner keys in owner custody; do not share seed phrases or signing keys with agents.
- Mailto sending is supported only when the site administrator has connected a real approved Mailto.si mailbox and Mailto HTTPS transport through the existing ModelGlue Agent Mail settings. Otherwise ModelGlue uses its explicitly configured WordPress notification transport; do not claim a Mailto message was sent without a delivery record.