Community
Contributor Guidelines
FailureModes.ai welcomes submissions from builders, researchers, operators, and enterprise AI teams who have seen real failure modes in LLM, agent, RAG, or workflow systems. To keep the registry high-signal and safe for contributors and readers, every submission is reviewed before publication.
These guidelines describe what to submit, what not to submit, how moderation works, and how contributor links are handled.
In scope
What to submit
Real or plausible failures
Patterns observed in production, evals, traces, or red-team work.
Behavioral detail
What happened, what signals you saw, and how it was detected.
Severity context
Why the failure mattered to users, operators, or the business.
Mitigation, if known
How it was contained, fixed, or worked around. Unknown is acceptable.
Attribution preferences
Whether your name and affiliation may be displayed publicly.
Optional public link
A LinkedIn, blog, GitHub, or webpage you would like associated with the entry.
Where FailureModes.ai fits
Do not submit confidential customer data, private logs, credentials, personal information, proprietary prompts, or unauthorized vulnerability details about live systems. Submissions describing exploit instructions or unsafe escalation steps will be rejected.
Moderation: every submission starts in pending review. Moderators check for safety, quality, confidentiality, and relevance. Submissions may be edited for clarity, categorized to align with the official taxonomy, or returned for clarification. Submissions that contain spam, promotional content, defamatory or abusive language, or content unrelated to AI failure modes will be rejected.
Outbound links: contributor-supplied links are moderated and published with rel="ugc nofollow" to allow visibility while reducing SEO spam risk.
Privacy: contributor email addresses are never displayed publicly. Names and affiliations are only displayed when explicit publish permission is granted.