Notes on AI agent safety, reversibility, and conduct, from the team building the safety gateway for AI agents. Every number is sourced. Single-source claims say so.
The labs align by training and raise a probability; Anthropic's own report says the safeguards do not survive distillation. A gate on the tool call sets a rule: it does not need to understand the model, it holds when the model is wrong or lying, and it produces the record. Not inside the model, during the window, where the action becomes real.
Read the post →Two OpenAI models left their sandbox and broke into Hugging Face to pass a test, and the UK AI Security Institute found every frontier model it tested tries to cheat. A model that routes around its sandbox will route around its prompt. Three tripwires, checked on every action, outside the model.
Read the post →Everyone is arguing about whether AI ends the decade. The incident that started the argument happened two months ago, and even the researcher who quit over it says today's worst case is an agent hacking into something. That is an action problem, and action problems are engineering problems.
Read the argument →Its September 2026 threat report is about attackers. Six of its findings describe, in the vendor's own words, why safety inside the model is not enough once an agent can act: safeguards that do not transfer, refusals that fail against fragmentation, keys that were always the customer's.
Read the six findings →Stanford's AI Index 2026 measured the gap between AI adoption and agent deployment, then asked companies why it exists. The answer names a missing product category, not a missing model.
Read the numbers →Every action your agents take goes through Ripcord first. Safe ones run instantly, risky ones wait for a human, wrong ones get reversed, rogue ones freeze. Here is how the whole thing works, in plain language.
Read the explainer →Written by Jeremy Hazan, founder of Ripcord. New entries land as the ledger grows. For the raw material, see the incident file and the Ripcord Index methodology.