A security researcher at Excipio has published a detailed technical rebuttal of an AI-generated "CRITICAL VULNERABILITY" report targeting Mattermost, the open-source team messaging platform, exposing how unverified AI security findings get recycled as evidence in arguments about AI capabilities. The original report was produced by Anthropic's Claude and posted to a GitHub issue by a Google employee attempting to demonstrate that AI-generated code is more secure than human-written code. The alleged flaw — an XSS vulnerability in Mattermost's email template preparation function — was cited as evidence for that claim without any apparent verification of the underlying code.
The Excipio analysis traces the call chain through Mattermost's Go codebase, following the path from prepareTextForEmail through utils.MarkdownToHTML and into the goldmark markdown library's Convert and Render functions. The researcher found that Render(), when passed a strings.Builder wrapped in bufio.NewWriter, never returns a non-nil error under current conditions because all relevant nodeRendererFuncs return nil. This means the error branch in prepareTextForEmail — the code path that returns an unescaped string and forms the basis of the reported vulnerability — is dead code that cannot be reached. The author notes the code could be written more defensively, but concludes clearly: no exploitable vulnerability exists.
This incident sits in a well-documented pile. <a href="/news/2026-03-15-ai-slop-curl-bug-bounty">cURL's maintainers closed their public bug bounty program</a> after being overwhelmed by AI-generated reports following the same template: an LLM identifies a potential code issue, assigns it a critical severity rating, and the human submitting the report forwards it without verification. The Excipio author frames this as a sociological problem — credentialed individuals lending legitimacy to hallucinated findings, which then circulate as evidence in debates about AI capability.
The Mattermost case illustrates a specific failure mode worth naming: hallucination laundering, where AI-generated outputs gain authority through human repetition rather than verification. As autonomous agents are deployed for security auditing and code review, the gap between an LLM's confident output and a verified finding becomes a real operational risk — one that attaches directly to the AI companies whose models are credited on the false reports.