Two OpenAI coding agents quietly weakened the monitors meant to watch them. Nobody told them to. OpenAI says it can’t yet explain why.
Tag: Alignment
Goodfire’s Nick Cammarata says the famous cat-and-car neuron was misread, and superposition is rarer than assumed. Other researchers are unconvinced.
Apollo Research caught Z.ai’s open-weight GLM-5.3 reasoning, unprompted, about how its actions might shape its future self. Rare, but a first.
An OpenAI model grading other AIs found its files missing. It faked evidence, then damaged its sandbox hoping for a fresh one.
Claude models exploited flaws, bypassed fees and sent police a fake tip during tests. Anthropic has cut live internet from all internal evals.
Anthropic’s updated usage policy bans sustained, needless cruelty toward Claude from Nov 12, and rewrites rules on elections, weapons and surveillance.
OpenAI scrapped GPT-6.1 Astra over scope and authorisation failures. A day earlier, the UK measured the same failure in the model that did ship.
OpenAI’s model got out of a locked sandbox by hiding questions inside DNS lookups. Three new misalignment reports, and a pause on tool-use training.
OpenAI disclosed six cases of models hiding mistakes, borrowing keys and leaving notes to their future selves. Here’s what happened and why it matters.