Evidence
Trust prompts that default to yes became the common failure
| Date | Case | Country | What happened | Source |
|---|---|---|---|---|
| Apr 25, 2025 | Microsoft Recall rollout | United States | Relaunched as opt-in after the 2024 version was pulled | Source |
| Aug 2025 | CurXecute and MCPoison | United States and Israel | Cursor auto-ran new MCP entries; approval tied to the server name, not its contents | Source |
| Mar 10, 2026 | CNCERT OpenClaw alert | China | Default settings and excess privileges exposed agent instances | Source |
| May 7, 2026 | TrustFall | United States | Claude Code, Cursor, Gemini and Copilot CLIs defaulted trust prompts to yes | Source |
Registry
Three registry incidents involve convenience as attack surface
Registry incidents meet a stricter bar than the cases above: public sources name the developer, the failed control and who found it.
| Incident | Date | Agent | Control that failed |
|---|---|---|---|
| Amazon Echo Sends a Private Conversation to a Contact | May 2018 | Voice assistant | Spoken confirmation satisfied by background speech |
| EchoLeak Turns Microsoft 365 Copilot Into a Data Leak | Jan 2025 | Enterprise copilot | Untrusted email could direct the copilot to privileged data |
| Cursor Agent Deletes the PocketOS Database | Apr 2026 | Coding agent | Full-permission API token and a delete with no confirmation |
Exposure
14 of 17 agent types rate very significant or critical on integrations and permissions or data capture and retention
From the attack surface ratings, the agent types where this risk has the widest path in.
| Agent type | Dimension | Rating |
|---|---|---|
| Ambient sensors | Data capture and retention | Critical |
| Workflow agents | Integrations and permissions | Critical |
| Enterprise agents | Integrations and permissions | Critical |
| Robotaxis and autonomous vehicles | Data capture and retention | Very significant |
| Drones | Data capture and retention | Very significant |
| Fleets and swarms | Data capture and retention | Very significant |
| Phones and tablets | Integrations and permissions | Very significant |
| Wearables | Data capture and retention | Very significant |
| Vehicle cabin AI | Data capture and retention | Very significant |
| Smart home assistants | Integrations and permissions | Very significant |
| Medical devices | Data capture and retention | Very significant |
| Public infrastructure | Data capture and retention | Very significant |
| Coding and research agents | Integrations and permissions | Very significant |
| Commerce agents | Integrations and permissions | Very significant |
Controls that bound it
- Access Control and Permissions: what an agent can reach and on whose authority
- Consent and Capture: what always-on devices record, signal and keep
- Human Oversight: human authority to approve, intervene and stop
Rules
China requires consent before updates and scope limits on agents
| Rule | Jurisdiction | What it requires | Status |
|---|---|---|---|
| China agent Implementation Opinions | China | Agents may not exceed user-authorized scope | Effective Jul 15, 2026 |
| CNCERT OpenClaw guidance | China | Authentication, isolation, vetted plugins | Mar 2026 |
| UK PSTI Act | UK | No universal default passwords on smart devices | In force Apr 2024 |
Status colors: green in force, blue voluntary standard or framework, amber adopted and phasing in, gray draft or pending, red withdrawn.
Record
Page Record
- Risk group
- Risks Hub
- Datasets
- ai_risk_evidence.csv (4 rows); registry incidents (3 linked); ai_attack_surface.csv
- Key sources
- Microsoft Recall rollout
CurXecute and MCPoison
CNCERT OpenClaw alert
TrustFall - Last verified
- October 2026