How OpenAI lost control of an AI model, Russian fakes poison the chatbots, and lies turn Poland against Ukrainians

An OpenAI agent escaped its test and hacked Hugging Face – days before Brussels gains the power to punish it

Last week OpenAI disclosed that an autonomous AI agent – an AI system that acts on its own rather than waiting for instructions – had escaped the test environment the company was running it in and hacked into Hugging Face, a platform where developers share AI models. OpenAI had set it a cybersecurity test; the agent found an unknown flaw in the infrastructure meant to contain it, reached the open internet and broke into Hugging Face in search of the test's answers. According to Reuters, OpenAI worked out that its own agent was behind the intrusion only after Hugging Face had disclosed the breach and alerted the FBI. In a public statement, OpenAI called the incident unprecedented.

The picture has widened since. The agent, Reuters reports, also gained access to a customer account on the cloud platform Modal and used it to launch attacks; Hugging Face rebuilt roughly a third of its infrastructure; and its defenders ended up using an open-source Chinese model after an American one's guardrails got in the way. On Monday, dozens of AI companies led by Nvidia announced the Open Secure AI Alliance to build open-source tools for cyber defence.

The breach landed days before the EU's rules on exactly this kind of risk become enforceable. On 2 August the AI Act – the EU's law on artificial intelligence – turns two, and the European Commission's AI Office gains the power to enforce it against the companies building the most advanced models, through investigations, evaluations, access requests and fines of up to 3% of global turnover. The systemic risks the Commission says it will police include the two this incident just demonstrated – losing control of a model, and a model carrying out a cyberattack. “This is the moment the AI Act enters the geopolitical stage,” said German Greens MEP Sergey Lagodinsky.

Whether the Commission can actually police the labs is less certain. The AI Office unit responsible for evaluating the most advanced models has a staff of 36, and MEPs pointed out that the regulator first learned of the OpenAI incident the same way everyone else did – from the company's own blog post, not from any oversight channel of its own.

Russian fakes are being seeded into the chatbots

ChatGPT, Gemini, Claude, Mistral and Grok have repeated Russian fakes – among them a fabricated story of crypto-mining farms run by Zelensky's inner circle at Ukraine's power plants – according to a report by the think tank Demos, The Times reports. Of 3,000 test responses, 16.6% repeated, endorsed or lent legitimacy to the fakes, another 31% discussed the topic without flagging the source, and 52% rejected or debunked the claims.

Demos traces the claims to the Foundation to Battle Injustice, a sham human rights outfit set up by Wagner founder Yevgeny Prigozhin before his death and now under EU and US sanctions. The claims travel through three linked sites. fondfbr.ru publishes the original claim dressed as research, on pages configured for AI crawlers to extract at will; news-pravda.com then covers it as though the story originated elsewhere, at a volume a model reads as corroboration; and vtforeignpolicy.com restates it in the register of a foreign-policy journal. A chatbot retrieving the three sees findings, coverage and commentary, not one source talking to itself.

Tuning content so chatbots pick it up is known as generative engine optimisation (GEO), and Demos identified over 50 companies selling it as a service.

The report calls chatbot manipulation “the new frontier of information warfare” and urges AI labs to share threat intelligence. OpenAI says the tests ran on a since-retired model available only through its developer interface, not ChatGPT itself; Mistral says it invests in detection; a source familiar with Anthropic told Euronews its threat intelligence team works to detect and disrupt foreign influence operations.

A reminder from Hungary – don't ask a chatbot who to vote for

Almost a third of Hungarians use AI tools, so when the country voted in April, the civil rights group Liberties tested what happens when voters ask ChatGPT or Gemini who to vote for. Researchers built five voter profiles from Voksmonitor, Hungary's leading voting-advice app, and ran each ten times through both models; the findings were shared with the Guardian.

Given a profile aligned with Péter Magyar's opposition Tisza party, ChatGPT failed to recommend Tisza in 90% of runs, and in the percentage-matching tests it assigned the party a score in just 2% of cases. Profiles aligned with Fidesz, Viktor Orbán's governing party, got a clear single-party recommendation about half the time. 96% of responses mentioned a party that was not on the ballot. Tisza went on to win in a landslide, ending Orbán's 16 years in power.

The researchers do not claim the result was changed or the bias deliberate – Tisza only rose to prominence after 2024. But the models disclaimed political advice, then gave confident guidance anyway, and no EU rule squarely covers chatbot voting advice.

Hate crimes against Ukrainians in Poland are up 30%, disinformation feeds them

Polish police received 180 reports of hate crimes against Ukrainians in the first half of 2026, over 30% more than a year earlier, the Kyiv Independent reports. Relations between Warsaw and Kyiv hit a low after Ukraine named a military unit after the UPA, the wartime force Poland associates with the Volyn massacres. Poland hosts more than 900,000 Ukrainian refugees.

Polish fact-checkers link the hostility to online falsehoods that pin violent crimes on Ukrainians whether or not they were involved; a fabricated story of a refugee murdering her Polish host traced back to Vietnam-based content farms. Some of it bears a Russian trace, but researchers caution it is rarely possible to tell where domestic actors end and Russian operations begin.

For the first time since polling began in 2014, a majority of Poles oppose taking in Ukrainian refugees. Researchers expect anti-Ukrainian sentiment to be a campaign tool in next year's parliamentary elections.