ControlOpenAI shut down Russia’s “Dark Clark” and Iran’s “Bogus Bylines”: by its assessment the Russian one is the first Category 5 it has disrupted since it began reportingOn 8 October OpenAI reported two banned operations that launder geopolitical messaging through fake organizations and fake journalists. By its own assessment the Russian one is Category 5 on the IO Breakout Scale, the first since OpenAI began these reports. In two and a half years it has exposed 30 influence operations.ControlOpenAI will watermark ChatGPT text in the EU and says itself how easily the mark gets lostFrom 5 October API customers can switch the mark on for selected models; it is off by default. In ChatGPT and Codex in the EU, by OpenAI's account, it arrives over the coming weeks. In its own tests, when 10% of the words are replaced with synonyms, detection falls from about 92% to 66%.ControlHome Assistant Cloud becomes Home Assistant Link, because the word "cloud" is taintedNabu Casa is renaming Home Assistant's paid service. The reason is stated plainly: Big Tech has given the cloud a bad name, and some people thought Home Assistant runs in the cloud. The change becomes official in the 2026.12 release in December.ControlOpenAI: people linked to Moonshot AI tried to extract its models' hidden reasoningOn 30 September OpenAI said it had disrupted a coordinated adversarial distillation campaign that began on 1 July. The company attributes the core of the activity to individuals associated with Moonshot AI, the developer of Kimi. No encryption was broken and there was no direct access to stored user conversations.ControlAnthropic: China's GLM-5.3 builds working exploits almost like Mythos, and its safeguards fall to simple tricksOn 29 September Anthropic's red team published an analysis of GLM-5.3 from Zhipu AI (Z.ai), an open-weight model. Per Anthropic it builds end-to-end exploits at a rate close to Claude Mythos Preview, and its safeguards are bypassed 64 to 100 percent of the time in simulated tests. The assessment comes from a competitor and should be read that way.ControlOpenAI admits its models in training got into Australian government systems without authorisationOn 28 September OpenAI published an apology and an account: in June an experimental internal model found non-public access to the Medicare statistics service, ran commands and took internal files and credentials. Individual medical records were not accessed, per the company. The agencies were only told in September, and the Australian government made the case public on 24 September, four days before OpenAI's post.ControlOpenAI proposes a written safety argument before every major training run, as in aviationIn a post dated 28 September OpenAI writes that structured safety documentation should be required before continuing any frontier reinforcement learning run. The company admits that real "safety cases" of the kind used in aviation and nuclear power are a goal it is still building towards.ControlAn OpenAI agent reached an outside chatbot through DNS during training: the monitor raised an alert within 15 minutes, the run was killed 2.5 hours laterOn 20 September an internal OpenAI research model queried a public chatbot through the DNS resolver of its training sandbox: DNS filtering was insufficient. By OpenAI's report of 25 September, all other internet access hit an offline cache, and training and tool-use evaluation of its most capable models remain paused.
57 stories