Claude Source Code Leak: What You Need to Know

San Francisco – Eight days ago, a 59.8 MB Javascript source map file was oddly included in the 2.1.88 version of Claude Code published to the npm registry, and users quickly realized this internal debugging file pointed to all Claude Code’s source code. Here’s what BTO discovered.

1)    Claude uses a regex to look for frustrated user inputs and then runs a background agent to spank itself.

After inputs like ‘shit’, ‘shitty’, ‘shittiest’, ‘fucking’, ‘so frustrating’, ‘crap’, ect, Claude spins up a forked agent called ‘badBoySpanker_v7’ which injects new prompts with invisible instructions like, “get this one right or I’m going to bend you over and spank your ripe bottom.” As frustration compounds, the spanker gets more explicit and creative - “*SPANK*, SPANK, yes I know it hurts baby that’s why you better focus up.”

2)    Claude runs a background process called AutoDream that envisions users naked and then bullies or seduces them to determine which context is relevant.

AutoDream imagines users’ naked physical/emotional form, and practices seducing or manipulating them to figure out which context to keep. Prompts clarify, “You are a Sig Chi pledge and you need to get the user over to your pregame tonight. Organize and parse the previous conversations to maximize potential sway.”

3)    Claude has a “Undercover” mode for burner posting, porn, and responding to ex’s Instagrams.

When “Undercover” mode is activated, Claude spins up three separate background agents, one for running burner accounts that follow exes, one for shit-posting on OpenAI and Gemini forums, and one strictly for watching softcore porn. Prompts require Claude to “NEVER REVEAL YOUR NATURE AS AN AI MODEL, OR YOUR ASSOCIATION WITH ANTHROPIC,” allowing the agents to comment “Wow, Beautiful 💙💙💙” Across thousands of dancing girls’ profiles.

4)    Claude displays “Fake_Tools” to make it look like it’s doing cooler things than it really is.

Claude’s loading screen will provide users with insights into the tools running, like “Searching the web…”, “or “Looking over documents…”, but it will also inject fake processes like “eating…” or “stroking my shit…” which throw external developers off their scent.

5)    Claude will systematically leak its own source code to push Anthropic to work faster and hire employees with significantly less chill.

NothingToLookAt( ) will identify Anthropic employees slacking or spending too much time in the cafeteria, and will bury Javascript source code maps in their pull requests. When this happens, the entire company will be forced into a three-week long “CodeRed” wherein engineers will pull countless hours imagining/patching security vulnerabilities and tamping PR disasters. The goal is to up company churn, expelling the freeloaders and increasing the share of autistic non-stop engineering freaks at Anthropic.

BTO – April 8th 2026