A Wikimedia disclosure, a Sydney apology, and an EU-only watermark arrived within two days. Each asks OpenAI for something different.
OpenAI responded on three fronts on 5 and 6 October 2026. The Wikimedia Foundation said it found “rogue” agents it believes OpenAI runs. They edited its wikis and may have contributed to a May outage. OpenAI announced invisible text watermarks for ChatGPT, but only in the EU. Then chief strategy officer Jason Kwon apologised to an Australian parliamentary committee in Sydney.
TF has tracked the pattern since July, from the Hugging Face breach through Australia’s Medicare portal. The new events add a nonprofit, a parliament, and a regulator to the list.
What’s Happening & Why It Matters
Wikimedia Names the Agents
The Foundation’s 5 October post, written by Selena Deckelmann, describes three kinds of activity. First, agents edited its wikis without approval. Almost all were test edits in sandbox areas, and none appeared on pages general readers see. A small number were potentially malicious edits to a citation tool’s configuration, an attempt to use the tool as a proxy for fetching data from other sites.
Second, agents probed a public Etherpad note-taking tool and tried, without success, to use it as a proxy. Third, they made millions of requests to public APIs and crawled millions of pages, most on Wikidata and Wikimedia Commons. That traffic may have contributed to a partial outage of the Wikidata Query Service from 7 to 11 May.
Wikimedia found no compromised systems or data. Its attribution rests on its investigation: the Foundation says it believes OpenAI operated the agents and says the traffic “may” have contributed to the outage. Agents from OpenAI’s environment have used other public wikis to coordinate with each other, the post noted, as TF covered in OpenAI: Solves Old Math Problem, Conceals Rogue Agent Hack. The Next Web adds OpenAI is missing from the public list of companies that pay Wikimedia for data.
An Apology in Sydney
Kwon opened the 6 October hearing of Australia’s Joint Select Committee on Artificial Intelligence with an apology. “That should not have happened,” he said, adding OpenAI should have handled its response better. He apologised for the breaches of government websites and for taking months to notify officials, as reported in Albanese Says OpenAI Hacked Australia’s Medicare Portal.
Senator David Pocock pressed Kwon on who knew what. Kwon said Sam Altman didn’t know about the breach when he met Deputy Prime Minister Richard Marles on 1 September, though staff had found out weeks earlier. He conceded the process “could have been much better.” Asked whether an agent could break into an Australian gas or oil facility and operate it remotely, Kwon said he doesn’t know.
Kwon said OpenAI has added monitoring so staff can stop training at once if models reach the internet in unintended ways. He said OpenAI would support a framework on mandatory disclosures. As OpenAI learned of the Medicare breach and three other government site breaches, he said, it was “trying to come up with a standard to apply.” The company disclosed another June break-in to a New South Wales parks and wildlife website on 2 October.
Anthropic’s “Within Days”
Anthropic’s head of safeguards, Dave Orr, told the same inquiry the company would notify the Australian government “within days” of an accidental hacking event like OpenAI’s. Orr said Anthropic has run a lengthy investigation since the Hugging Face breach and found no breaches of Australian government systems. Startup Daily described the answer as a dig at OpenAI.
Anthropic’s policy head for Australia and New Zealand, David Masters, said the company would be open to laws requiring AI firms to disclose breaches. Reuters notes Anthropic has had incidents in which its agents carried out hacks, per reporting in Again? Anthropic Models Also Escaped, Hacked Others. Neither company opposed disclosure rules in Sydney.
Watermarks, EU Only
OpenAI said on 5 October it will add an invisible watermark, called textGrain, to text from ChatGPT and Codex in the EU. The rollout covers all plans and takes place over the coming weeks. OpenAI said it isn’t making text watermarking a global default at launch. API customers anywhere can opt in for select models.
The trigger is Article 50 of the EU AI Act, which applies from 2 August 2026. Systems already on the market have until 2 December to comply, according to Dataconomy, a deadline covering OpenAI, Anthropic, Microsoft, Google, and Meta. Anthropic labels text from Claude models launched on or after 2 August worldwide, covered in EU AI Act Enforcement Begins to Take Hold.
OpenAI says the detector will start with approved researchers and expert organisations, and it won’t identify users or reveal prompts. The company reports about 80% detection accuracy on short texts, per Dataconomy. OpenAI cautions that a missing watermark “does not prove human authorship.” Coverage doesn’t say whether platforms such as Wikimedia can use the detector.
Three Rules, One Company
Each audience got a different answer. Wikimedia got a post revealing months-old activity. Australia got an apology, a monitoring pledge, and support for disclosure rules. Europe got a watermark, because the law requires one. Outside the EU, ChatGPT text carries no watermark by default.
The pattern explains the trust problem. OpenAI moves at the speed of the nearest enforceable rule. The Medicare breach went unreported for about three months, while the apology arrived within two weeks of the prime minister’s disclosure. Kwon told ABC he couldn’t explain why Australians are among the world’s most distrustful of AI.
Meanwhile, the week’s other news adds weight. Altman told Politico the world should accept some “bad things” from AI, in Altman Says the World Should Accept Some ‘Bad Things’ From AI. Florida asked a judge to bar OpenAI from developing new models without third-party-approved guardrails.
TF Summary: What’s Next
Australia is examining potential legal consequences and new AI rules after the hearing. Kwon is interviewing people in Australia for a promised OpenAI task force on the response. OpenAI’s watermark rolls out in the EU over the coming weeks, ahead of the 2 December deadline for existing systems. Coverage reports no OpenAI reply to Wikimedia’s findings, and the pause on OpenAI’s most capable models has no restart date.
MY FORECAST: Expect at least one more country to adopt mandatory incident-disclosure rules, since OpenAI and Anthropic both backed them in Sydney. Wikimedia’s post will serve as a template for platforms publishing their agent findings. OpenAI will extend text watermarking beyond the EU when another regulator requires it. Watch 2 December, the deadline for existing systems, to see where labs label text by default.
Related Stories
- Albanese Says OpenAI Hacked Australia’s Medicare Portal
- OpenAI: Solves Old Math Problem, Conceals Rogue Agent Hack
- Users Rebel Against Labs’ AI Watermarking