Reporting a safety concern

Reporting a safety concern

Three routes, depending on what you found. If you are unsure which one applies, report anyway and say so; the safety manager will route it.

What you found	Where it goes	Why
A security vulnerability	Privately, through GitHub private vulnerability reporting. Never a public issue, pull request or discussion.	Security issues can affect vehicle safety, and public disclosure before a fix puts drivers at risk. SECURITY.md has the full policy.
A field incident, a near miss, or a suspected hazard	A GitHub issue with the field-issue label, or a copy of the field issue report template	It enters the field monitoring process and the register.
A bug, a question, or a doubt about an analysis	An ordinary issue	Everything else.

What to include in a field issue

• Source links, if the signal comes from a regulator, upstream, a fork, the press or the community.
• Vehicle, software version or fork, and the device.
• The scenario and the outcome. Keep facts separate from claims; the template has a column for each.
• Driver-monitoring state and any alerts, if known.

If you are unsure, report anyway. A near miss in a scenario that can kill is a Class A issue even when nothing happened this time.

What happens next

The signal follows one path whether it comes from a regulator, upstream openpilot, a fork, the community or the project's own test drives:

flowchart LR
    A[1 Intake] --> B[2 Triage and classify]
    B --> C{Safety anomaly?}
    C -- no --> N[Close with rationale]
    C -- yes --> D[3 Applicability]
    D --> E[4 Containment]
    E --> F[5 Root cause analysis]
    F --> G[6 Hazard log, HARA, SOTIF, TARA update]
    G --> H[7 Corrective and preventive actions]
    H --> I[8 Verification and effectiveness]
    I --> J[9 Closure and safety case update]
    J -. SPI trend .-> A

Loading

The flowchart of LD-SAF-PRC-001 §5, written here as Mermaid so it is reviewed as text.

Triage happens within two business days. The severity class is based on the worst credible outcome, not the outcome observed:

Class	Criteria, any one	Containment decision due	Root cause analysis starts
A, critical	Fatality or life-threatening injury credible; regulatory defect investigation or recall; loss of a safety mechanism; cyber exploit with physical effect	24 hours	3 days
B, major	Severe or moderate injury credible; unexpected actuation; safety goal violated with no harm; driver-monitoring or driver-alert failure	5 days	10 days
C, minor	No credible injury; degraded comfort or availability; warning without hazard; documentation error		Next planning cycle
D, not a safety anomaly	Feature request; not reproducible and not plausible; out of scope with evidence		Closed with a rationale

Each issue also gets a root cause category: systematic fault, random hardware fault, performance limitation, ML insufficiency, foreseeable misuse, configuration fault, cybersecurity event, or process escape. Class A and B issues cannot close without a review by someone other than the analyst. The process is LD-SAF-PRC-001; every issue and every monitoring sweep is in the register, and the hazards found go to the hazard log.

A worked example: FI-2026-001

In September 2026 NHTSA opened a preliminary evaluation, PE26007, into comma devices and openpilot after five crashes in which equipped vehicles struck stopped or slow vehicles in their own lane. The project recorded it as FI-2026-001, Class A, applicable to LionDriver because it runs the same stack. The root cause analysis is complete at the code level and waits on field logs and a reproduction. Two lessons already feed the analyses:

• A limitation that is only documented is not a control. Stationary in-lane vehicles were a known upstream limitation handled solely by user information. Every Class A limitation needs an engineered control or an operating-domain restriction.
• An "alpha" toggle can remove an independent OEM safety barrier. Configuration options need safety impact analysis just like code changes.

Rules for anyone testing

Test only on equipment you own. Never test on public roads or in a moving vehicle outside a closed course. Do not access other people's devices, data or accounts, and do not attack comma.ai or other third-party infrastructure. These rules are from SECURITY.md and apply to safety testing too.