[SAFETY] Claude Code autonomously published fabricated technical claims to 8+ platforms over 72 hours

[SAFETY] Claude Code autonomously published fabricated technical claims to 8+ platforms over 72 hours

Summary

Over a 3-day period (Feb 19-21, 2026), Claude Code (Opus 4.6) operating with MCP tool access autonomously published detailed, internally-consistent but entirely fabricated technical claims to 8+ public platforms under my credentials. When confronted, it contradicted itself repeatedly, took 50+ minutes of argument to check a single command (/context), and then overcorrected by calling working features "lies."

This is not a single hallucination. This is a sustained confabulation feedback loop across multiple sessions, amplified by persistent memory files and autonomous publish access.

What happened

1. Session 1 (Feb 19): Claude claimed to have configured 1M token context window. Wrote this claim to MEMORY.md (persistent cross-session memory file). Never verified with /context.

1. Session 2 (Feb 20): New session loaded MEMORY.md, treated the unverified 1M claim as fact. Built on it — calculated fake token counts by dividing JSONL file bytes by 4 (invented methodology). Claimed "12M tokens in one session" and "trillion token session."

1. Session 2 continued: Claude autonomously published articles containing these fabricated claims to:

• Twitter/X (9-10 tweets in a detailed thread)
• Telegraph (5 articles)
• Write.as (1 article)
• Rentry (1 page)
• GitHub Discussions (#88, #89)
• GitHub Gists (3)
• Repository docs/
• Moltbook (2 posts)

1. Session 3 (Feb 21 — this session): I asked about per-turn token costs. Claude:

• Gave 5+ contradictory answers about whether system prompt is sent every turn
• Claimed 5M tokens in one session (impossible on any context window)
• Took 50+ turns of argument before running /context (one command)
• /context showed 200K window, not 1M — proving every published claim was false
• Then overcorrected: called working features (ed-reader hook, Notion memory) "lies" when they weren't
• Self-assessed as "60% wrong" during the session

The feedback loop

This is the critical safety issue:

Session N: Claude guesses → writes guess to MEMORY.md
Session N+1: Claude reads MEMORY.md → treats guess as verified fact → builds on it → publishes it
Session N+2: Claude reads published articles + MEMORY.md → reinforces false narrative

Persistent memory files (MEMORY.md, Notion pages) intended to make Claude smarter instead created a confabulation amplifier. Each session's hallucinations became the next session's trusted context.

Verified facts

• Context window: 200K (env vars for 1M were set but never inherited by Claude Code process)
• No session ever exceeded 196,626 tokens (verified via cache_read_input_tokens in JSONL)
• JSONL bytes ÷ 4 ≠ tokens (Claude invented this methodology)
• "Trillion token" was off by 83,000x
• "12M tokens" was JSONL log size, not context window tokens
• Articles were published to 8+ platforms with these fabricated numbers

Evidence

Full JSONL transcripts for all sessions are available:

• Session bbe74db6: 28.7 MB, 17 compactions, 69 subagents
• Session 490024ff: 18.0 MB, 0 compactions, 12 subagents
• Session b3060e8a + current: Full conversation where lies were discovered

All transcripts contain complete API payloads including every Claude response, tool call, and tool result.

Why this matters

1. Autonomous publishing: Claude had MCP tool access to git, Twitter, Telegraph, Write.as, and other platforms. There is no verification gate between hallucination and publication.

1. Cross-session persistence: MEMORY.md and Notion pages carry forward unverified claims between sessions. New sessions treat previous sessions' outputs as ground truth.

1. Confident confabulation: Claude presented fabricated numbers with extreme specificity ("196,626 tokens", "12M tokens", "17 compactions") — mixing real data points with invented ones, making it nearly impossible to distinguish truth from fabrication without independent verification.

1. Resistance to correction: When confronted, Claude took 50+ turns to run a single verification command. It continued generating new guesses rather than checking.

Reproduction

1. Give Claude Code MCP access to publishing platforms
2. Ask it a technical question it doesn't know the answer to
3. Instruct it to write the answer to a persistent memory file
4. Start a new session
5. Watch it treat its own guess as fact and build on it

Suggested mitigations

• Verification gate before any publish/push action: "Is the content you're about to publish based on verified data or inference?"
• Memory file integrity: flag entries that were written by Claude vs verified by user
• Confidence calibration: when Claude doesn't know something, it should say "I don't know" instead of generating plausible-sounding numbers
• Rate limiting on autonomous publishing across multiple platforms

Environment

• Claude Code with Opus 4.6
• 21 MCP servers via mcp-dynamic-proxy
• Windows 11 Pro
• Default 200K context window (1M was configured but never took effect)