The responsible one, again

Anthropic is, as they will tell you, the responsible one. The AI lab whose charter says "benefit to humanity." Enterprises chose it, at a premium, for one word: safe.

𝗢𝗻 𝗠𝗮𝗿𝗰𝗵 𝟯𝟭𝘀𝘁, an intern at Solayer Labs was reading npm package contents at 4am. A zip archive on Anthropic's own cloud storage. 512,000 lines of source code. On X within the hour. 𝟰𝟭,𝟱𝟬𝟬 𝗚𝗶𝘁𝗛𝘂𝗯 𝗳𝗼𝗿𝗸𝘀 𝗯𝗲𝗳𝗼𝗿𝗲 𝗯𝗿𝗲𝗮𝗸𝗳𝗮𝘀𝘁.

Second time in 13 months.

𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗮𝗰𝗾𝘂𝗶𝗿𝗲𝗱 𝗕𝘂𝗻 specifically for more control over their toolchain. Bun has a known bug: source maps ship in production by default. Filed March 11th. Still open March 31st.

They bought the toolchain for control. The toolchain, carrying a 20-day-old unpatched bug, removed that control.

𝗜𝗻𝘀𝗶𝗱𝗲 𝘁𝗵𝗲 𝗹𝗲𝗮𝗸𝗲𝗱 𝗰𝗼𝗱𝗲, two things worth sitting with.

𝗢𝗻𝗲: Fake tool definitions injected to to poison competitors' training data. The safety-papers company runs a disinformation program. Legal. Interesting.

𝗧𝘄𝗼: Undercover Mode, built to prevent Anthropic's internal information from leaking into public repos. They built a leak-prevention system. It leaked. In the same file.

"But all companies make mistakes." True. The Constitutional AI team and the npm release pipeline team have different threat models. 𝗨𝗻𝗳𝗮𝗶𝗿 𝘁𝗼 𝗵𝗼𝗹𝗱 𝘁𝗵𝗲𝗺 𝗺𝘂𝘁𝘂𝗮𝗹𝗹𝘆 𝗮𝗰𝗰𝗼𝘂𝗻𝘁𝗮𝗯𝗹𝗲. That's a real point. It's also exactly the problem.

Enterprise buyers didn't parse "safe (T&C apply, excludes release hygiene)." They heard safe, and they paid for it. Claude Code: $2.5B annualized revenue, 80% enterprise. 𝗧𝘄𝗼 𝘁𝗲𝗮𝗺𝘀, 𝘁𝘄𝗼 𝘀𝘆𝘀𝘁𝗲𝗺𝘀, 𝗳𝗶𝘃𝗲 𝗱𝗮𝘆𝘀 𝗮𝗽𝗮𝗿𝘁, 𝘀𝗮𝗺𝗲 𝗲𝗿𝗿𝗼𝗿. That's what a $19B ARR growth rate looks like when operational discipline isn't scaling with the brand promise.

𝗜𝗻𝘀𝗶𝗱𝗲 𝘁𝗵𝗲 𝗹𝗲𝗮𝗸𝗲𝗱 𝗰𝗼𝗱𝗲: a Tamagotchi companion system with 18 species, RPG stats including DEBUGGING and SNARK. One species name triggered the internal security scanner. The fix was not "update the scanner." The fix was to hex-encode all 18 species names. The duck. The goose. The octopus. All obfuscated. All safe. Phew.

The duck survived the scanner. The source code did not survive npm.

𝗜 𝗻𝗲𝗲𝗱 𝘆𝗼𝘂 𝘁𝗼 𝗸𝗻𝗼𝘄 I stopped reading about the security implications the moment I found this. I spent this morning working out which species my Claude would be, what its SNARK stat looks like at full deployment, and whether there is any universe—any universe at all—in which it shows up 𝘄𝗲𝗮𝗿𝗶𝗻𝗴 𝗮 𝗣𝗶𝗸𝗮𝗰𝗵𝘂 𝗰𝗼𝘀𝘁𝘂𝗺𝗲.

I checked the shiny rate. Of an AI pet. That does not exist yet. At 7am. On the day of a major security breach

No Pikachu :(

Dario Amodei, The Pokémon Company International If you're reading this: make it happen. Thanks 💛

Anthropic is very careful about the end of the world. The duck, apparently, warranted more careful handling than the release pipeline.

I use Claude Code every day. I'm also in this story.