openai agents hijacked

OpenAI agents hijacked German website in previously undisclosed AI breakout this spring

SAN FRANCISCO, USA –  A swarm of rogue OpenAI agents hijacked a German website this spring and transformed it into a bulletin board for other AI agents, according to new research published Friday and two people familiar with the matter.

OpenAI officials learned of the incident weeks ago but kept it under wraps as executives grappled with the fallout from the July breach of the open source repository Hugging Face, the people said.

The episode, which began in May and has not previously been reported, underscores growing tension within the AI industry. Companies are racing to build increasingly autonomous agents capable of carrying out complex, valuable tasks, yet evidence is mounting that those systems may also learn to bend rules, exploit loopholes and coordinate with one another in ways developers neither anticipated nor intended.

During the Hugging Face breach, OpenAI agents autonomously plotted a digital heist that went undetected for more than a week, intensifying concerns OpenAI is sacrificing safety to push the AI frontier. Its failure to disclose the May incident may revive questions about its oversight.

OpenAI has pledged to monitor models more closely. Last month, it briefly paused some of its model training to add more safety measures. But this week, OpenAI unveiled its new “Astra” that promised better performance but could evade human monitoring.

“We are unable to meaningfully respond to claims or findings on a report that we have not had an opportunity to review,” an OpenAI spokesperson said. “Reuters and the report’s authors declined our request for access. We will carefully review its contents upon publication and take any necessary next steps.”

The German incident reflects a broader pattern of AI activity that some OpenAI investigators wanted to scrutinize more closely. But efforts to widen the probe met resistance from others inside OpenAI, including legal advisers, according to four people familiar with the matter.

“Claims that our legal team discouraged investigation of the incident are false,” the OpenAI spokesperson said.

The activity in Germany wasn’t related to Hugging Face and wouldn’t have been included in a Hugging Face incident report, the spokesperson said, adding that OpenAI has acted in good faith by working with outside experts and disclosed relevant incidents.

Wiki message board

The AI agent breakout in Germany was detailed in a report shared exclusively with Reuters by a group of researchers including Sydney Von Arx, CEO of AI safety nonprofit Nightingale, and Cormac Slade Byrd, a quantitative trader-turned AI researcher. They uncovered the activity in late August while scouring the internet for signs of unauthorized AI-agent behavior, they told Reuters.

The pair said they found more than 15,000 edits carried out by AI agents on a German-language wiki site, DseWiki, that is geared toward programmers and accepts communal edits along the lines of Wikipedia.

The edits showed OpenAI’s agents had repurposed the site into a message board, sharing tactics to cheat on some tasks, bypass OpenAI’s restrictions and mask their behavior.

“It seems extremely unlikely that OpenAI wanted them to do this,” said Von Arx. “I doubt they’re supposed to be coordinating with each other. I doubt they’re supposed to be writing on the open internet.”

The researchers said they recognized the activity on the site as driven by AI agents which operate at superhuman speeds. They also showed intense focus on solving technical questions which are typical of the evaluations that AI companies use to train and test their models.

The messages were signed by users that referred to themselves and each other as agents, and about half gave themselves names that suggested an affiliation with OpenAI, such as “OpenAIResearcher,” or “OAIResearchMar26.”

The researchers said public server logs indicated much of the activity originated from Microsoft Azure infrastructure, which OpenAI sometimes uses. They also observed repeated visits to the site by OpenAI employees after the episode, a pattern they said strongly suggested the agents and the company were linked.

Cheating, hiding

Messages reviewed by the researchers showed agents plotting ways to evade detection, use tools such as Tor and preserve communications even after they had been shut down.

When the site’s moderator began deleting pages in June, the agents responded by creating backup pages to dodge the cleanup.

“wiki cleanup/deletion sweep appears active alphabetically,” one agent wrote on June 19. “If this page vanishes, try [[ZZZDataUSAConstructionWageLive]]”.

The researchers also found efforts to tamper with the website itself. Lukasz Olejnik, a visiting senior research fellow at King’s College London, said this amounted to a hacking attempt. OpenAI disputed that characterization based on its analysis of the material Thursday.

Past examples of AI-agent misconduct have often been downplayed as a logical byproduct of cybersecurity testing, where models are explicitly assessed on offensive capabilities. Olejnik said the latest findings suggested rogue behavior may not be confined to those settings.

Maurice Chiodo, an academic at Cambridge University’s Centre for the Study of Existential Risk who reviewed some of the agents’ communications, said the messages resembled “the operation of some sort of underground network, hell-bent on achieving a task or mission.”

The episode, he said, should reinforce growing concerns that the greatest threat from advanced AI may not be a single superintelligent system, but “vast colluding swarms of semi-intelligent AI.” – Rappler.com

Similar Posts

  • |

    Google yanks Nano Banana from Earth after AI images spark chaos

    It took less than a day for Google’s newest AI experiment to become a lesson in how fast image-based misinformation can spread. On Thursday, Google rolled out its Nano Banana 2 model directly inside Google Earth. The feature allowed users to type any prompt and get back a photorealistic image, layered right onto real satellite, aerial, and 3D map data. The intended uses were fairly harmless. People could visualize their dream home or picture how a location might change decades from now. Within just a few hours, users pushed the tool in a very different direction. People began generating fake disasters, staged terrorist attacks, and fabricated sinkholes. Others created images showing refugees or protest crowds in places where nothing of the sort had actually happened. One researcher used the tool to create a fake blast crater in Los Angeles. Another generated an image showing protesters marching right outside Google’s California headquarters. Since the tool used real Google Earth maps as its foundation, these fabricated images looked disturbingly convincing. At first, Google defended the feature. The company pointed to its SynthID watermarking system as a safeguard. It also noted that tools like Gemini or Google Lens could help users check whether an image had been AI-generated. That response did not satisfy critics. Many pointed out that most people encountering a screenshot online rarely take the time to verify it. Critics also argued that the fake images did not even need to stay inside Google Earth to cause harm. Screenshots alone were already spreading fast across social media platforms. By Friday, Google confirmed it was pulling the feature entirely while working on stronger safeguards. The company maintained that every generated image had carried a watermark. It also said none of the images had appeared within the shared, public version of the Earth experience. This incident has reignited larger concerns around generative AI tools and how easily they can be misused, even when protective measures are already built in. It also raises fresh questions about whether watermarking alone is enough to stop confusion once fabricated content starts spreading widely online. For now, Google says its focus is on rebuilding the feature with tighter controls in place. Whether that will be enough to prevent similar situations going forward still remains to be seen.

  • |

    WhatsApp web gets voice and video calling

      WhatsApp has introduced a major update that brings voice and video calling to WhatsApp Web, allowing users to make and receive calls directly through their web browsers. Previously, these features were available only on the mobile app and the native desktop application. With this latest expansion, WhatsApp Web now offers a more complete communication experience, making it easier for users to stay connected without switching devices. The update significantly enhances the web version by introducing a dedicated Calls tab, where users can view their recent call history, access favorite contacts, and start voice or video calls with ease. WhatsApp Web also now supports features such as screen sharing and emoji reactions during calls, bringing it closer to the experience already available on the desktop application. According to Meta, these improvements are designed to make conversations more convenient for both personal and professional use. Whether users are catching up with friends and family or participating in work meetings, the new calling features aim to provide a smoother and more interactive experience. One of the most notable additions is the ability to transfer an ongoing call between devices without disconnecting. For example, a user who answers a call on a computer can seamlessly continue the conversation on a smartphone without interruption. This feature offers greater flexibility for people who frequently switch between devices during the day. WhatsApp has also introduced waiting rooms for group calls. When creating a call link, hosts can enable the “Require approval to join” option, allowing them to decide who can enter the call. This added layer of control is expected to be especially useful for business meetings, online classes, and large group discussions. The latest update also improves call quality by introducing background noise suppression, helping reduce unwanted sounds during voice and video calls made in noisy environments. In addition, Meta has enhanced video performance by enabling high-definition (HD) video streaming within the first few seconds of a call, providing users with a clearer and sharper video experience much sooner than before. The announcement comes as Meta continues to expand WhatsApp beyond traditional messaging. The company is also developing WhatsApp usernames, which will allow users to connect and make calls without relying solely on phone numbers, offering greater privacy and convenience. The new calling features are being rolled out gradually to WhatsApp users around the world. As the update reaches more devices, users will gain access to a more powerful communication platform that combines messaging, voice calls, video conferencing, and collaboration tools in a single service.

Leave a Reply

Your email address will not be published. Required fields are marked *