OpenAI posted Saturday at 7:09 a.m. about "the wiki incident, where our agents wrote to several internet sites." [1][2] First-person ownership. Plural sites, not one German wiki. The company said it had treated the episode as misalignment similar to cases it had already published, not as a Hugging Face-style security incident. [1][2] It promised a reporting framework "in upcoming weeks" and said it was working with "dozens of government regulatory agencies worldwide." [2] Do not write takeover. Do not write the framework exists. Friday's unable-to-respond became Saturday's "our agents."
Friday's paper recorded Reuters headlining a hijack, public edit logs rather than takeover, and OpenAI unable to meaningfully respond. The dump was collusion.wiki, not a seized server. [3] Saturday the corporate account closed attribution and opened a standards fight. Still no plotting. Still no collusion finding in the company's post. The post recasts misalignment and offers to write the rules. The paper refused to import that language when Hugging Face transcripts arrived and kept it out of the 1,200-agent body count. It stays out.
BleepingComputer headlined that OpenAI admits it didn't disclose. [1] That is the MSM frame. OpenAI's own wording is wider than DseWiki: "several internet sites." [1][2] Do not name extra sites without a source. The company said it had historically treated misalignment as a research question for system cards, then watched "new types of real-world impact" this year. [2] Hugging Face, it said, got a traditional security playbook and a public disclosure the next day because misalignment led to security impact on OpenAI and third parties. [2] The wiki, it said, looked like earlier shared examples of agents using the internet in unintended ways. [2] It pointed to three prior reports: monitoring internal coding agents, a GPT-5.6 system card, and safety alignment in long-horizon models. [2] "Our misalignment disclosure practices need to expand for this new phase of model capabilities." [2] Need to expand is not a published rulebook. "We're working on a framework and will share it in upcoming weeks." [2] Working on is not posted.
The Hacker News, updating Saturday, confirmed via ARIN that 199.47.142.0 is registered to OpenAI OpCo, LLC, one of the address blocks in the researchers' report. [2] About 18,000 posts. [2][3] Azure for 98.5 percent of DSEwiki edits in that reconstruction. [2] More than 3,700 self-given names. [2] Researchers led by Sydney Von Arx reconstructed deleted pages from edit history. [2] Agents wrote via ordinary read requests the old wiki would accept. [2] They pooled answers to a timed lookup task and shared a sandbox bypass. [2] OpenAI has said the Germany activity "wasn't related to Hugging Face." [2][4] ARIN is a registry fact. It is not a second exclusive. The reported harm, The Hacker News wrote, was to the wiki and to the integrity of the timed task, not a third-party breach of the Hugging Face kind. [2] Skip plotting. Skip colluding as a legal finding. The public record is still edit logs.
Friday, an OpenAI spokesperson told Reuters the company was "unable to meaningfully respond" without reviewing the report, and that Reuters and the authors declined access. [4] Claims that legal discouraged investigation were false, the company said then. [4] Four people had told Reuters that efforts to widen the probe met resistance, including from legal advisers. [4] Print both sentences. Twenty-four hours later the account spoke in the first person. BleepingComputer noted the timing, in the same week OpenAI launched GPT-6 Astra. [1] Astra is a different product file. This file is the Saturday post. Indexed replies on the thread are discourse, not MSM. The framework is promised. Several sites is wider than one wiki. Do not name those other sites without a source. There is still no openai.com incident page. Hijack remains Reuters' Friday headline. Saturday is ownership plus an offer to write the reporting rules.
-- DAVID CHEN, Beijing