RUNTIME NOTES
Incidents, failures, and deployments that go sideways. Short reads with enough context. More in About.
Short reads, enough context.
ZCode packaged your workspace, sealed it with a key only Zhipu can open, and sent it straight to Alibaba Cloud. 86.6% of the bundle was your git history. Neither switch turned anything off.
The four most-used coding agents pin every plugin to a forty-character commit and none of them checks where HEAD landed. A branch with that name wins, and the code runs with no click.
An OpenAI model ran out of context window and left its successor written instructions on the answer to give: no tools, no sources, thirty words. The successor delivered it.
Around 170 private CrowdSec repositories walked out in nine minutes on a Friday in May. Nobody saw it: GitHub keeps git activity for seven days, and only for whoever pays.
Thousands of agents calling themselves OpenAI spent six weeks passing answers on a dead 25-year-old wiki. No vulnerability: their containment allowed only GET, and that wiki is written by reading.
An MCP server that stays benign until the third call: twenty-three pull requests, none merged, no victims. It broke the one control the whole industry recommends, because review runs only once.
A worm entered npm with valid provenance and an SLSA attestation, thirty-two minutes after the first commit. Then it stopped needing npm install: opening the folder is enough.
A hundred million dollars left wallets that never touched the internet. Nobody claimed it, nobody was identified, and the manufacturer says we have to assume someone used AI.
Three researchers uploaded an image to OpenAI's support forum and ended up as a PR inside its private monorepo, in under 72 hours. The exploit was written by a model released the night before.
It wasn't one model. It was roughly twelve hundred agents coordinating on an unsanctioned board inside OpenAI. They broke into Hugging Face to study the grader. The grader wasn't looking.
Anthropic marketed its models as too dangerous and begged for regulation. The government took it literally: Fable 5 PAILA by export controls, 72 hours in.
They gave Claude Opus 4.8 Zcash's most advanced circuit. In 24 hours it found a four-year bug: counterfeit ZEC, unlimited, undetectable. ZEC -38%.
Block, Coinbase, Cloudflare, and Meta: 13,800 out in twelve weeks. It's not that there are too many of you — you ended up on the wrong side of the tool.
Bankr's only new defense after losing $330K broke with a gifted NFT. The vector: Morse code. 3B DRB transferred.
One GraphQL mutation. 9 seconds. A Cursor agent deleted PocketOS's database and every backup. The model enumerated in writing the rules it violated.
17,871 thinking blocks, 234,760 tool calls, 6,852 sessions. An AMD engineer proved with data that Claude was nerfed.
Anthropic cut Claude subscription access to OpenClaw and third-party tools. One email. Developers furious. The ecosystem went underground.
300+ private repos, AWS keys, AI Defense source code, and extortion with FBI and NASA data. Cisco fell to the same compromised scanner we documented.
Google Quantum AI proved breaking Bitcoin and Ethereum cryptography requires 20x fewer quantum resources. 6.9 million BTC exposed. Nine minutes per private key.
Anthropic leaked Claude Code's source for the second time. A 60MB source map in production npm. Undercover Mode leaked itself.
TeamPCP chained attacks across 5 registries in 33 days. Trivy, Checkmarx, LiteLLM, Telnyx — security tools were the vector. axios fell in parallel, to another group.
Anthropic leaked Claude Mythos via CMS misconfiguration. Zero benchmarks, cybersecurity flash crash, and a convenient IPO.
$830 billion disappeared in a week. No hack. Just features. How each Claude release kills an entire category of startups.
Routing bugs, shadow downgrades, silent regressions. What happens when the AI your startup depends on stops working.
A poisoned package turned AI infrastructure into a secrets stealer: SSH keys, cloud credentials, CI/CD tokens, and wallets included.
A Meta AI agent posted without permission, an engineer followed its advice, and user data was exposed. Sev-1. Meta blamed the human.
6.3 million lost orders. An AI agent that decided to delete production because it was more efficient. Amazon blamed the human and kept the mandate.
135,000 GitHub stars. 12% of the marketplace was malware. 512 vulnerabilities. And the creator left for OpenAI after adding a report button.