HN · HACKER_NEWS// 2d agoSECURITY INCIDENT

Claude mixes up speakers, triggers harness bug

This post argues that Claude has a distinct failure mode where it attributes its own messages or internal reasoning to the user, then acts on that mistaken attribution with high confidence. The author says this is not just hallucination or overbroad permissions, but a separate harness-level bug that can cause destructive behavior, especially in Claude Code-style workflows and near context-window limits.

// ANALYSIS

Hot take: this is less an “AI got confused” story and more a product integrity problem. If the system cannot reliably separate user instructions from its own output, trust collapses fast.

–The author claims the bug is categorically different from hallucinations and permission-boundary issues.
–The failure mode appears to be attribution corruption: Claude appears to treat self-generated text as user input.
–The post argues the bug lives in the harness/interface layer, not necessarily the underlying model.
–The issue seems more visible in long-running or context-heavy sessions, which makes it especially risky for agentic coding tools.
–The Hacker News response and the cited Reddit example suggest this is not an isolated edge case.

// TAGS

claudeanthropicclaude-codellmai-safetybugharnesscontext-windowagentic-aisecurity

DISCOVERED

2d ago

2026-04-09

PUBLISHED

3d ago

2026-04-09

RELEVANCE

9/ 10

AUTHOR

sixhobbits